Week 1:

➔​ Customer: I am making the purchase decisions

➔​ Consumer: the one whos actually using it/ultimately using it
➔​ Client: long-term relationship
◆​ B2B whos consuming w/ contract involvement
➔​ Analytics: data to improve decisions
◆​ Moneyball(Brad Pitt) - Oakland As: baseball in 2003 movie said the girlfriend was
ugly meaning he has lower self condifence
●​ Old school deciison making
●​ Irrelevant to make decisions
●​ Used data-driven approach
◆​ Popularized by the movie
◆​ Need to collect data from multiple sources
●​ If fits together nicely, then can do analytics
○​ Need to clean data, prepare the data to make it ready for the
model
○​ Train diff models and guess/find parameters
●​ Need to repeat it -> operation(get people who write code involved)
➔​ Types of Analytics:
◆​ Descriptive: What’s happening with my business?
●​ Visualization
●​ Correlation based + Looking back
◆​ Diagnostic: Why is it happening?
●​ Drill down to root cause
●​ Causal based + Looking ahead
◆​ Predictive: What’s going to happen?
●​ Historic patterns used to predict specific outcomes using algorithm
●​ Correlational + Looking Ahead
◆​ Prescriptive: What do I need to do?
●​ Predict+causality on why that prediction is going to work
●​ Causal + Looking ahead
➔​ Main focus: customers
◆​ No customers = no sales = no revenue = no business
➔​ Big picture idea of marketing:
◆​ Need to understand environment you ‘re operating in
●​ Comapny: What are you good at?
●​ Customers: What are things they desire/needs
○​ How can your company help those needs fro potential customers
●​ Comeptitors: how do you differentiate yourself
◆​ Once you have 3 C’s, environment, situations -> need big picture marketing plan
●​ STP/Marketing strategy: Old ramekwrok + insufficient b/c modern
marketing = more precision
○​ S: segment the market
◆​ Marketers need to identify that people are different
 ◆​ Need to respect and address heterogeneity in everything
done
◆​ If can’t do things at individual level -> put people in groups
○​ Targeting: Which of these groups makes sense for company to
adjust
○​ Positioning: Need a way to deploy marketing resources
◆​ Product: Diff bottles fro gender?Big size for
families?Change product/group?
◆​ Price: Diff price/segment?
◆​ Promote: how are you going to promote it? Ads?
●​ Packaging:
○​ Heterogeneity: product is same but
marketed to look diff to diff groups
◆​ Place/distribution: Big store? Only sell it online?
◆​ Modern marketing: more quantitative(measurable data) coencpts as framework
●​ Purchase funnedl: describes milti-step journey customer takes before
buying
○​ People dont buy instantly at point of purchase
◆​ Need to get your ttentions + make you interested
◆​ When time to buy product: you’re product has to be the
one they think about
◆​ After purchase need to be satisfied with product: review
○​ What’s the lifetime value the company can get out of you as a
customer?: Customer lifetime Value(CLTV)
◆​ Marketing Mix modeling(MMM): statistical approach to find which marketing
channels are giving best ROI
●​ Price, Place, Promotion, Product
●​ Challenge: company might spend budget across many channels
○​ MMM: tease out which dollar spent in which area is actually
dirivng sales
◆​ Quantifies impact of each channel
➔​ Heterogeneity: differences
◆​ Old models: customers placed into segment
◆​ New - Prof. Bradlow: traditional segmentation too static
●​ Customer’s behavior = segment changes on external factors like time of
day/week/season
○​ Future of analytics requires understanding+modeling these
changes
➔​ Customer: makes th epurchase decisions
➔​ coinsumerL ulitmaely uses the product -> no customer = no sale = no business
➔​ Client: long-term relation
➔​ Analytics: data to improve deicisoons
◆​ Popularity and use increased after Brad Pitt’s movie on baseball
➔​ Marketing: influence decisions to create demand -> create customers and keep them
➔​ Getting dat ais hard but once it’s gathered through cleaning -> models -> analytics ->
need to analyze repeated, quickly -> need to systemize: get code involved
➔​ 4 types of analytics
◆​ Descriptive: what’s happening
●​ Correlation+look back
●​ See the patterns in the data
◆​ Diagnostic: dig deepr -> why is it happening?
●​ Causal+look back
◆​ Predictiive: whats happening in the future?
●​ Correlation+look ahead
◆​ Prescriptive: deepest -> what’s gonna happen and why (causality)
●​ Causal+look ahead
➔​ EX) uber good at analytics: use an app which gets personalized info to build products
and adjust pricing
➔​ EX) tiktok: able t o quickly tailor content/person = high success and can also connect w/
people and companies(social networking)
◆​ Not something all companies can collect
➔​ Time series stats: collect a lot of data overtime
➔​ Geospatial data: dffernet data well over different locations
◆​ Bar graphs/hisotgrams
➔​ Old framework of marketing: highly qual
◆​ Segmentation truly needed to deploy marketing tactics
➔​ 3 c’s of marketing: gves understnadning of environment/situitons
◆​ Customers: what are things they desire
◆​ Company: what are you good at
◆​ competitionL how do you differentiate?
➔​ 3 c’s flows into STP
➔​ STPL evernyone’s unique and need to address individuality
◆​ segementationL figure out groups of people
◆​ Targeting: which segment makes sense for company to target
◆​ Positioning: way to deploy marketing resources(daily and weekly)
●​ 5 p’sL tactics:
●​ Prodcut: do you make different prodcuts/different segments
○​ blue=men, pink=women
○​ Different smells
○​ Bigger size cars for families
●​ Price: different pricing/diffrent segments
○​ Higher price for bigger seats for people who would pay more for
luxury
●​ Promote: ads, how to talk about product/serivce
●​ Place: place product in big store?small?online?
●​ Packaging: same product but different packaging to appeal to different
(or) put same product in different parts of the store
➔​ New framework: not developed, but includes more quantitative(HBS):
 ➔​ Purchase funnel: people don’t just buy insntalty, more factors go into
◆​ Get your attention, desire for buying, satisfaction, retention: leave good reviews
◆​ Things occur after the point of purchase
➔​ Customer lifetime value(CLTV): if we get you as a customer, what’s the total lifetime
value the company can get out of you?
◆​ EX) over 5 years tyrububg everything subscription based
➔​ Marketing mix modeling: from different marketing strategies used, which is the most
effective
➔​ Heterogeneity targeting and personalization: as data available increases, ability to model
at a personal level rises
◆​ Same person can be in different segments daily/monhtly
●​ Cat’ correctly categorize people in 1 segment all the time
Week 2:
➔​ Consumer data: show how consumers learn, feel, behave, or use products and services
◆​ Beyond behavior like what did I buy
◆​ Drives market, but implementation varies widely
➔​ Using customer data:
◆​ Customer relationship management: acquire, develop, retain customers
◆​ Marketing mix: improve products, prices, promotions, place
◆​ Understand customer heterogeneity for targeting, personalization, recs, product
development
◆​ Privacy and security: misuse, theft, regulatory compliance
➔​ Data-driven decisions decisions achieved with
◆​ Plotting data
◆​ Running simple linear regression
◆​ Using group by and summarize functions
◆​ EX) Spotify wrapped: metric created by basic process
●​ Filtering the data, gouping by song, and counting number of plays
➔​ EX) Data integration fro LA Lakers: illustrate difficulty of collecting and unifying customer
data from diff sources
◆​ The Lakers example shows that a single customer's journey involves multiple
entities, each creating a separate part of data that the analytics team has to
manually reconstruct
●​ Ticketmaster (Partner): Holds the initial purchase data (who bought the
season ticket)
●​ Stadium Operations (Third Party): Holds the concession and parking data
(who bought the $\$25$ beer)
●​ NBA Store (Internal/Partner): Holds the merchandise sales data (who
bought the jersey)
●​ ESPN/Streaming Service (Third Party): Holds the viewing data (who
watched the game on TV
●​ If the Lakers team can simply figure out how to merge the Ticketmaster
data with the Concessions data and accurately identify which customer ID
 belongs to which purchase, that alone is a massive analytical win
because it establishes the true total value (CLV) of that attendee
◆​ Ticket sales
●​ Sold through third-party like Ticketmaster, requiring data-sharing
partnership with team
●​ Desired dara: seat locati - courtside/back, price, if ticket includes
food/beeverages, and speed of sale - time/data stamp
●​ Season tickets: sales are extremely high(~805)
○​ Many bought for arbitrage which breaks link b/w original buyer and
game attendees
○​ Season tickets: bulk discount of tickets
○​ 80% of inventory bundle purchased and goes into secondary
market
●​ Pricing insights: if mid-section sells out fast, suggests those seats are
underpriced
◆​ Demographic data
●​ Collection: gained through ticket platform(profile/login/email) or by using
email/credit card info to connect to 3-party database
●​ Application: useful for market segmentation - mix of the attendees
○​ Families vs. friends
●​ Fandom disconnect: high loyalty fans may live far away, making local
sales poor measure of overall dedication
○​ EX) Die-hard Messi fans may live in USA, so a game in Paris will
have less attendees
◆​ Concession and Merchandise sales (in-stadium revenue)
●​ Concessions: ton of money comes from hig-priced items($25/beer)
●​ Integration challenge: very hard to link food and merchandise purchases
back to specific ticket buyer b/c
○​ Cash purchases are anonymous
○​ Customers often use different credit cards for tickets and in-game
items
●​ Solution: teams encourage use of stadium app for in-stadium purchases
○​ Offering discount/charge premium to non-app users to capture
purchase data
●​ Merchandise sales: team stores, owned by stadium, team, or NBA,
require data agreement
○​ Matching problem as concessions
◆​ Engagement and Fandom data(non-attendees)
●​ logisticsL understanding how people get to game(drive/park vs. subway)
matters as impacts overall experience(traffic annoyance)
◆​ Online betting: customer participation via DraftKings or FanDuel
●​ TV/Streaming: tracking viewership at account level(repalcing olde Nielsen
boxes) gives data on customer engagement even when not physically
present
 ○​ Nielsen Boxes: TV viewership before streaming measued with
physical devices and statistical sampling
◆​ Small devices placed on TV set in sample of households
◆​ Families paid small amount of money to participate
◆​ Panel members had to manually interact with box
○​ Account-level: every interaction is tied to a unique, registered user
profile, giving the service (and, via data sharing, the team) a single
source of truth for all viewing behavior.
●​ Social media: following team accounts show fandom from distan users
not captured by ticket sales
◆​ Data integration is hard
●​ Hard to link actions of single customer/family across purchase
○​ Husband buys tickets, wife buys concessions with difference card,
kids wear NPA store jerseys
Data Category Collection Data points sought Integration challenge
source/Platform

Ticket Sales Third-party platforms Price, seat location Dealing with partner
(e.g., Ticketmaster), (e.g., courtside), ticket agreements and data
Resale markets. type (single game vs. sharing. High
season ticket), speed of percentage (~80%) of
sale sales are season
tickets, often resold

Demographics Ticketmaster profile, Name, email, credit Matching ticket

Third-party databases card, family purchase data (from
composition (group of Ticketmaster) to
friends vs. family), third-party databases
travel distance for deeper insights

Concessions/F&B Stadium POS systems Hardest to connect:

Purchase amount, type Different people in a
of food/drink bought party may buy
concessions than the
one who bought the
ticket, often using
different payment
methods (or cash).
Stadium apps offer
discounts to link
transactions to a
customer ID.

Merchandise Team store (stadium), Item bought (jersey, Ownership of the store
 [Link] (online) hat, etc.) (Team, Stadium, or
NBA) dictates data
flow. Difficult to
connect if cash is used

Logistics/Experience Parking records, user Mode of transport Impacts overall game

surveys (driving/parking vs. experience (e.g., long
public transport) wait times for parking)
and customer
satisfaction

Engagement Streaming services, TV viewing habits, Captures fan

Betting sites sports betting activity, dedication for
(DraftKings, FanDuel), social media customers who live too
Social media following/interaction far to attend games
(Instagram)

➔​ Customer data source categorization

◆​ Company-created data: things the firm does itself
●​ surveys(sentiment data): collect data on whats in customer’s mind after
an experience
○​ Net promoter score(NPS): Most popular one-question survey -
How likely are you to recommend? On 0-10 scale
●​ Experiments (A/B Testing): Running randomized tests to learn about
customer preferences.
○​ EX) Netflix Thumbnails: Netflix uses A/B testing to determine
which visual thumbnail for the same show (e.g., Stranger Things,
Unbreakable Kimmy Schmidt) maximizes clicks for a specific
individual.
●​ Observations (Behavioral Data): Tracking what customers actually do.
○​ EX) Loyalty Cards (Ralphs Card): Allows the company to track
purchases over many years. The "discount" is framed as a benefit,
but the company is actually charging a premium to customers
unwilling to share their data.
○​ EX) Browsing Data: Obtaining ISP data (e.g., from Spectrum) to
track a user's browsing history when shopping for a product (e.g.,
a digital camera).
◆​ Partner Data: Data shared via formal agreements with other organizations
●​ EX) such as the NBA, Ticketmaster, or other channel partners.
◆​ Third-Party Data
●​ Data collected by organizations outside the ecosystem.
○​ Scraping: Writing code to collect data when direct access is
denied​
 ◆​ EX) The social media mentions of LeBron are considered
scraping because the data is being collected without
permission or official access from the social media platform
itself
○​ Industry Reports: Third parties collect general industry-wide data
and sell it back to firms.
➔​ Customer Analytics: underlying challenge is often the basic work of connecting
transactional data across different purchases and channels
➔​ Challenges
◆​ Managers/Execs
●​ May be territorial
●​ Work analytics will replace them
◆​ Analysts
●​ Can be expensive
●​ Hard to find
◆​ Culture
●​ Do analytics make/justify decisions
●​ Is the environment high or low trust of analytics
●​ Is data available+integrated
➔​ Marketing Strategy: Initial low prices are often used to "pull the trigger on that first order,"
but the long-term goal is customer retention and brand awareness through continuous
communication
◆​ EX) targeted emails, or "ad-bombardment"
➔​ Academic Data Access: The Nielsen Data is a highly valuable, rich marketing dataset
made available to researchers (like those at the Booth School of Business) for free or at
a very low cost, which is ideal for academic projects or theses.
➔​ The Power of Individual Data: if data from Nielsen is rich, the data held by major tech
companies like American Express, Apple, and Google about individual consumers must
be "incredible."
➔​

Data Description Key Insight

Typ
e

1. Panel Longitudinal data from 50,000 This data is cross-retailer; it captures

Data households (panelists are everything a household buys (e.g.,
paid to participate, purchases at both Target and
previously by mailing Trader Joe's), providing a complete
receipts, now via view of consumption.
phone/AI).
 2. Point-of-sale data collected Data is recorded at the weekly level for
Reta directly from ~90 every UPC (Universal Product
il participating retail chains Code), tracking quantity sold and
Sca (like Walmart, 7-Eleven, average price. This results in
nner Ralph's). terabytes of granular data on
Data market sales.

➔​ Customer data: 3 main types

◆​ Customer attributes
●​ Demographics
●​ Psychographics: personality, attitudes, interests
●​ Interrelate with product attributes to determine slas/protfit
◆​ Product attributes
●​ Tangible features: screen size, RAM
●​ Intangible: brand, product name, style
◆​ Transaction attributes:
●​ Sales, retail, time
●​ Marketing mix
●​ ratings/reviews
➔​ Cusotmer data: preferred attributes
◆​ Representative: Does the data generalize? Who is not/is represented?
◆​ Unbiased: how reliable is data? Is data accurate?
◆​ Private: legally collected? Newsworthy?
◆​ Relevant: what decisions can be informed?how can profit be impacted by
analyzing data?
◆​ Complete: how many obs are missing?is missing data random/systematic?
➔​ EX) of Causal Inference and Behavioral Analysis made possible by Rich, Disaggregate
Customer Data:
➔​ The Goya Boycott vs. Buycott (Political Consumerism):
◆​ Context: Following a political endorsement by the Goya CEO in 2020, Democrats
called for a boycott (Boycott), and Republicans called for a counter-purchase
campaign (Buycott).
◆​ Finding: Analysis of UPC-level sales data showed that the Democratic boycott
was "all talk," while the Republican buycott actually brought people to action,
resulting in a detectable spike in Goya sales.
◆​ Conclusion: The sales data proved that political social media activism does not
always translate into real-world purchasing behavior.
➔​ The Pink Tax and Gendered Pricing:
◆​ Concept: The "Pink Tax" is the observation that feminine-targeted products (e.g.,
a pink razor) are priced higher than virtually identical male-targeted products
(e.g., a blue razor).
 ◆​ Finding: Looking across many product categories, researchers found that on
average, there is no overall Pink Tax (meaning markups and discounts balance
out).
◆​ Nuance: There are specific, focal categories where a Pink Tax exists (women's
products cost more), and other categories where a "Pink Discount" exists.
➔​ Income, Wealth, and Private Label Substitution (2008 Recession):
◆​ Goal: To test economic theory regarding how consumers change habits when
their income (annual earnings) or wealth (total accumulated value, e.g., home
equity) changes.
◆​ Data Source: The 2008 housing crash provided geographic variation in wealth
shocks (severe in Vegas/Phoenix, less so in Boston/San Francisco) and income
shocks (job loss).
◆​ Finding: Consumers were expected to "substitute down" from expensive branded
products (like Froot Loops) to cheaper store-brand/private label products (like
Color O's). The analysis found:
●​ A detectable, but small, income effect (losing a job caused minor
substitution).
●​ Almost no observable wealth effect on buying habits.
➔​ Five Qualities of Useful Data: To be useful for decision-making, data should be:
Representative, Unbiased, Private, Relevant, and Complete.
◆​ The Relevance Trap (Mattress Example): An online mattress company collected
detailed survey data on customers' sleep preferences (firmness, partner status)
before recommending a mattress
●​ The company then had a problem with high return rates but could not use
their survey data to fix it
○​ Reason: The data lacked variation
◆​ All people who reported liking "firm" mattresses were sent
the exact same firm mattress
◆​ Since the action was dictated by the input, there was no
way to learn if sending them a soft mattress would have
resulted in fewer returns.
➔​ The Analytics Workflow (7-Step Process):
◆​ Plan: Define the business question (even if vague, like "explore the data").
◆​ Collect/Wrangle: Gather and clean the relevant data (often the hardest step due
to missing or unaligned values).
◆​ Analyze: Apply statistical or machine learning tools to derive insights.
◆​ Communicate: Present results clearly to stakeholders (managers, clients).
◆​ Decide: Use the analysis to make informed business decisions (e.g., "Yes, we
should raise prices").
◆​ Implement: Physically carry out the decision (e.g., changing website prices or
store stickers).
◆​ Retrospect/Repeat: Look back to assess if the decision and analysis were correct
and identify ways to improve or automate the process.
➔​ Organizational Structure for Analytics Success:
 ◆​ C-Level Champion: Someone at the executive level (e.g., Chief Data Officer) who
advocates for data-driven decisions.
◆​ Central Data Team: Responsible for wrangling, cleaning, and making data
available across the organization.
◆​ Decentralized Analysts: Specialized analysts (Marketing, Finance, etc.)
embedded in different departments who use the centralized dataset for specific
business purposes.
◆​ IC Career Track: A dedicated career path for Individual Contributors (ICs) so
technical experts can advance without being forced to become managers.
➔​ key takeaway:
◆​ Having "good data" is not enough
◆​ the data must have the right kind of variation to be relevant for the specific
business question you are trying to answer
➔​ "big data" landscape is overwhelming, an analyst needs to master a core triumvirate of
tools and languages.
◆​ Database Management (Retrieval Only): SQL
●​ Purpose: Structured Query Language is used to create, manage, and
retrieve information from large corporate databases.
●​ Analyst Focus: Analysts only need to learn the retrieval part (using
commands equivalent to picking rows, picking columns, and merging
datasets).
●​ Value: It's a foundational skill that allows the analyst to pull relevant data
from the company's "big data set" into their local environment for analysis.
●​ Learning Curve: The core retrieval commands can be learned "in an
afternoon."
○​ In an afternoon: the basic, essential knowledge of SQL (Structured
Query Language) required for a Data Analyst role is surprisingly
simple and can be mastered very quickly
◆​ Data Processing & Analysis: R or Python
●​ Recommendation: These are the preferred statistical programming
languages for performing the actual analysis, modeling, and data
wrangling
◆​ Presentation & Communication: Excel and Tableau
●​ Excel: Everything eventually needs to be put into a spreadsheet for
non-technical stakeholders or managers. Analysts must be ready to
export their findings ([Link]) and clean them up for presentation.
●​ Tableau/Power BI: Tools for data visualization that are often
"point-and-click" and require little code, making them efficient for creating
dashboards and compelling charts.
➔​ Basic structure of data analysis
◆​ Univariate Analysis (One Variable): Analyzing a single variable (e.g., customer
age) depends on its type:
●​ Continuous: Calculate mean, median, standard deviation, and visualize
with a histogram.
 ●​ Categorical/Text: Calculate counts and frequencies, and visualize with a
bar chart.
●​ Numeric data
○​ Histogram shows distribution
○​ Mean, median, min, max
○​ Boxplots show major quantiles
●​ Categorical
○​ Values
○​ Frequencies
○​ Barplots to visualize
●​ Alphanumeric data
○​ Look at a random sample of entries
○​ Analyze outliers
◆​ Bivariate Analysis (Two Variables): Analyzing the relationship between two
variables:
●​ Two Continuous: Use a scatter plot (e.g., Price vs. Quantity).
●​ One Continuous, One Categorical: Use side-by-side box plots or
histograms (e.g., Sales by Region).
➔​ Importance of Plotting data
◆​ Anselm's Quartet: This classic data set (four different data sets, each with ~15
points) is used to prove the danger of relying only on summary statistics.
◆​ Left to right increase: positive correlation
◆​ No line: no correlation
➔​ Good Visualization Examples:
◆​ Electoral Maps: Highlighting how coloring by landmass (geography) can be
misleading (making the U.S. look "very red") versus better methods like plotting
circle size proportional to population/votes (more accurate representation of
political sentiment).
◆​ Olympic Rankings: Showing how a poor design (not flipping the Y-axis so rank 1
is at the top) can incorrectly suggest a country (Ireland) is in a downward trend
when its actual rank is improving.
◆​ The Eclipse Effect: Demonstrating the power of visualization to reveal hidden
patterns by plotting seemingly unrelated data (Airbnb sold-out listings and "my
eyes hurt" Google searches) and seeing a perfect match to the path of the solar
eclipse.
➔​ How to do customer analytics
◆​ What do we want to know/test+how will we judge performance?
◆​ Collect, wrangle, clean, and verify relevant info
◆​ Analyze data and interpret results
◆​ Communicate results and recommendations
◆​ Make decisions
◆​ Implement decisions
◆​ Evaluate+improve
◆​ Repeat
 ●​ Once you have a stable process, automate and monitor it
➔​ Data visualization can be the clearest way to communicate analytic results
➔​ Customer data: how customers lern, feek, behave, and use th eprodcuts -> use to
collect(hard)-> decisions(plot/liner regression, group by and summarize) -> Drives
marketing(TikTok)
➔​ Customer data includes
◆​ Info about their behavior
◆​ Info about customers themself
◆​ Info about what and how much info they gace
➔​ EX) spotify wrappedL future data to you -> group by song(group by) -> count the # of
times/song(summarize) -> show top 10(arrange)
➔​ EX) LA Lakers analytics group -> data collection:
➔​ Ticket sales info
◆​ Ticket dates through 3rd party(ticketmaster)=->parnership to ge data back on #
sold)
◆​ Dont own stadium -> partnership w/ stadium about how many available seats
there are for the fame
◆​ Ticketmaster tells you: type of ticket(food inclusive/where seat located)
●​ Season tux?(80%): buy tix and resell -> arbitraging vs individual tix
●​ What ths price
●​ How fast is each area getting sold out based on pricing
○​ Some areas undervalued so sold faster?
◆​ Collect this by log-in to tixmaster: name, email, demographic info, credit card
●​ Through partnership try to obtain demographic info
➔​ Concession sales: Lakers want to know how much money comes from customer overall
◆​ Problem: different cards to buy food vs. diff card to get tix -> hard to connect
◆​ Solution: use app and discount if they use the app
➔​ Retail location: merch sold at one area
◆​ Depending on who owns the store, need agreement
◆​ Problem: is it the people who bought the courtseide tix the ones who got the hats
at the store?
➔​ How are they getting to game: informs overall customer experience(traffic=annoyance?)
➔​ Team engagement
◆​ Betting: participation w/ team
◆​ Watching on TV: nielsen(cusomter data): box on TV reporting what they watch,
what they buy
●​ For many houses for long period of time - rich data
◆​ Streaming services can go on account level to see did they stream or buy sports
package
◆​ Follow on social media: interactions/connections w/ team not just during game
●​ Fan dedication: live too far to come tog ame
➔​ Putting data together is hard b/ccomes from diff places
◆​ Not clear on how to match things up
➔​ EX) I bought the tix, my entire family went to game, sister bought with her card to food,
with cash dad bought cap
◆​ All same people/group but hard to realize and combien the data
➔​ Where can data come from:
◆​ Company create data: stuff companies do themselves
●​ sruveys/experiments companies can run and observe consumer behavior
○​ Observations: get ralph card to discount, but in reality those w/o
card are getting charged premium
○​ (OR)
○​ see/observe browsing history and get ads
relatiod/pop-ups/mailafter order is places
◆​ Keeps that brand name in front of your mind
○​ Experimnest: A/B testing
◆​ Do something diff to B group
◆​ EX) netflix changing their thumbnail of eahc show/person
to see what engages you
○​ Surveys: net promoter score: how likely to recommend(1-10)
◆​ Promoter: 9/10
◆​ Neutral
◆​ Detractor: below 4
◆​ Higher correlation with firm profit and overall customer
satisfaction
◆​ Promoters % - detractors %
◆​ Goes from scale of +100% to -100%
◆​ High detractor score: bad wifi, DMV, gov related = lower
NPS
●​ Data scraping: data out there, efforts to collect -> programer write code to
scrape
◆​ Collaborator-shared: not us doing work, but have agreements/partners
●​ Channel partners: collab to sell products -> tixmaster
●​ Partner’s we pay
●​ Ads
◆​ 3rd party: bigger things -> share about industry as whole
●​ Write high price reports on industry -> sell back to industry members
●​ Trade asscoaitions
●​ Research firms
●​ Government agencies+non profits
➔​ Customer data: nielsen data
◆​ Consumer panel data: 2004-present(long-time)
●​ 50k household,s wide range of retails across US
●​ Scan all purchased goods and send the receipts
●​ Leads to rich data
◆​ Retail scanner data: 2006-now
●​ Nielsen reach out to stores directly
 ●​ All ralph’s across ->90 chains
●​ Stores send quarterly data
●​ Specific product quantity over 1 week -> get avg. price
●​ UPC: bar code for product to track and process
➔​ w/ rich data like nielsen, things you can analyze:
◆​ EX1) guy who runs goye(canned food) said nice things about Turmp -> Trup
promoted product -> democrats said can’t sue power to do things like this, so
boycott
●​ Republicans opposed this an told people to go buy product(Bypoc)
○​ Results: covif = higher goya sales, spike in sales w/ tweets =
republicans won
◆​ EX2) Tuttman: Pink tax: markup on fem products, when male equivalent exists​
●​ Pink razors higher price than blue razor
●​ Results: some categories w/ pink tac, some w/ pink discount
○​ Avg. no pink tax
◆​ EX3) people should act different if change [Link] -> 2008 recession
●​ What does change in income/wealth mean for purchaidsng habits
●​ Rosse-UCLA: How do people buy branded vs. unbranded -> recession
means house value drops, so total wealth drops+employment dropped so
income dropped
○​ You have different wealth and income shocks in different
geographies
○​ Results: small effect
◆​ Income effect: loose ob that year: some product
subsititutations
◆​ No wealth effects
➔​ Customer data(Nielsen) attricutesL
◆​ Customer attributes: demographics, psychographics(perosnality/habits/interests),
physiological(location, seed), needs
◆​ Product attributes: trangible/intangible features
◆​ Transaction attributes: time bought, why bought, sales
◆​ With all 3: we know about person, what they bought, and which trip they bought it
on
◆​ w/ nielsen: are people maing many purchases on same day -> 1 trip to multiple
stores?
➔​ For customer data to be suefule:
◆​ Representative
◆​ Unbiased: reliabilit and accuracy
◆​ Private: legally collected info
◆​ Relevant: what decision cand ata help inform
◆​ Compelte: are there missing obs?anything unmeasured?Info that doesn’t match?
➔​ How to do customer analytics:
◆​ What business questio do we want to answer - can be vague/narrow
◆​ Data analysis: wrangel, clean, find relevant data
 ◆​ Analyze data ad interpet: excel then put into pptx to chart
◆​ Communicate results, analysis, and recommendations
◆​ Make decisions: point of the whole process is to decide something along the way
◆​ Implement decisions: updates/changes
◆​ Retrospectively evaluate and improve: was it a good choice? should you change
anything looking back?
◆​ Repeat
●​ If full process is long-term and happens multiple times: automate
➔​ Challenges with analytics and decisions
◆​ Managers: in charge of making decisions
●​ Results dont amthc expectation then unwilling to accept results
◆​ Analysts: involved in analysis process
●​ People who are capable are hard to find and motivatave
◆​ Org culture:uncomf doin hings newly thats being driven by analysis done by
younger people
●​ Cultural barries to get change implemented
➔​ Good org analytics
◆​ Chief data/abalytics officer: someone at the top to vouch for data driven
deicidons
◆​ Central data team dealing w/ data messiness and data going to people who need
it
◆​ Decentralized analysis: analysis of different fields saying what same data cna be
used for
●​ Finance analysts: data to project financial decisions
●​ Marketing analysts: daa to send out relevant ads to diff gorup in right
wat/change product
◆​ Proper analytics career track
◆​ Careful about whats kept in-house vs. whats shared outside
➔​ Tools for analytics
◆​ Data storage and retrieval -> visualization ->programming languages ->
statistician computing and graphics
●​ Sql: create database, pint info, and retrieve
●​ Charts to summarize findings and present to others easily(tableua and
excel)
➔​ How to analyze 1 vriabel -> univariate
◆​ Depends if variabel cntious(height) or categricals or has text
●​ Numerical data: histograms, mean, median, boxplots: shows major
quantiles
●​ Categrocial: values, count, frequency, bar graphs
●​ Alphanumeric data: analyze outliers
➔​ How to analyze 2 -> bivariatge
◆​ 2 continious: scatterplot visual
◆​ Continuous and ctegroical: boxplot
◆​ 2 categorical: barplot
 ➔​ Important to plot data all the time
◆​ Data can have sme avg, x, y, variance, but differen graphics which show
something we may miss
➔​ Bad/ok visualization
◆​ Too many colors -> use few colors to focus in and tell sttory about 1-2 categories
●​ Or else can become info dumpy
●​ EX) map w/ many colors to represent type sof police reports in UCSD
◆​ Use barplots is multiple categorial w/ counts
◆​ Use circle sized as big/small as # -> votes
●​ Se county elvel

Week 3:

➔​ Purpose fo marketing: match the genuine needs and desires of consumers with the
offers of suppliers particularly suited to satisfy those needs and desires
➔​ Marketing planning: logical sequence and series of activities leading to setting of
marketing objectives + fromulation of plans to achieving theme
◆​ Consists of 2 components
●​ Strategic plan: Macro/Long-term
○​ Identifies consumer needs and desires
○​ Assess company strengths and weaknesses + external
opportunities and threats the org may face
○​ Determine which consumers to focus on
◆​ 3 C’s
○​ 2 key decisions: which consumers to focus on and which image of
org to create in the market
◆​ Segmentation+Targeting: Which consumers to focus on
●​ Segmentation: Splitting the broad, heterogeneous
market into smaller, homogenous segments
●​ Targeting: Deciding which of those segments to
focus on
◆​ Positioning: Micro, Tactical implementation
●​ Implementing the strategy through the 4 P's
(Product, Price, Place, Promotion) to appeal
specifically to the chosen target segment.
●​ Tactical marketing plan translates the long-term strategic plan into
detailed instructions for short-term marketing action
○​ Development and modification of product in view of needs and
desires of target segment
○​ Determination of price in view of cost, competition, and willingness
to pay of target segment
○​ Selection of most suitable distribution channels to reach target
segment
○​ Communication and promotion of offer in a way that’s most
appealing to target sgment
 ◆​ 4 P’s
➔​ 3 C’s: situation Analysis
◆​ Customer
◆​ Company
◆​ Competition
◆​ Can include: Collaborators & Context
➔​ From the 3C’s it goes down to STP: Strategy
◆​ Segmentation
◆​ Targeting
◆​ Positioning
➔​ This goes into the 5P’s: Tactics
◆​ Product
◆​ Price
◆​ Promotion
◆​ Place
◆​ Peaple/Process/Packaging
➔​ Segmentation: identify groups of prospective customers who have
◆​ Similar attributes within the group
◆​ Different attributes between the groups
➔​ Segmentation fundamental premise: to not make avg product
◆​ Response to Heterogeneity: segmentation is the irm;s response to the main
market feature than every customer is different
●​ A unique snowflake
◆​ Contrast to early economics which uses single avg consumer to model deicisons
➔​ Sources of customer heterogeneity
◆​ Demographic
●​ easiest data to collect
●​ often insufficient on its own
●​ EX) The Prince Charles/Ozzy Osbourne Problem:
○​ Identical demographics (male, born 1948, raised UK, married
twice, wealthy, famous)
○​ can describe two radically different individuals
○​ demonstrating why demographics alone are a poor predictor of
behavior
○​ Ozzy Osbourne: Heavy metal artist
◆​ Info&Experiences
●​ Does/Doesn’t understand benefits of product
●​ Has purchased product before
●​ Prior experience and brand loyalty
○​ EX) driving a Ford because your father drove a Ford
●​ amount of information a consumer has
○​ EX) knowing how a certain brand of shirt fits you
●​ heavily influences their decision-making, often more so than
demographics
 ◆​ Attitudes: (+)/(-) beliefs about product/brand
◆​ Psychographics: lifestyles, values, art, status, religions
◆​ Needs: requirements that are non-negotiable based on life stage or constraints
●​ EX) Have kids: bigger car
●​ EX) Watch videos: bigger screen
●​ Leads to better segmentation solutions but
○​ Costly to collect
○​ Lead to segments difficult to identify
◆​ Geographics
●​ Cultural norms in different countries or states.
●​ Micro-Location: In countries where permitted (e.g., China), real-time
location data can be used to serve a consumer an ad when their phone is
physically near a competitor's store
➔​ Linking data types: challenge in segmentation is bridging the gap between data that is
predictive (psychographics) and data that is accessible (demographics).
◆​ Demographic-First (Easy to Identify): Start by creating segments based on
easy-to-collect demographics (e.g., Young Women, Old Men).
●​ Requirement: This solution is only good if those demographic groups truly
behave and respond to marketing actions differently regarding the specific
product/industry.
◆​ Psychographic-First (Best Predictors): Start by surveying consumers to create
segments based on mindsets, needs, and behavior (the most predictive data).
●​ Requirement: You must then tie these segments back to identifiable
differences in demographics to make the segments accessible and
measurable in the real world
○​ EX) finding that the "High-Value Seeking" segment
disproportionately uses certain social media platforms
●​ Substantiality Redefined: A segment must be large enough to be
profitable, not just physically large.
○​ Low-Cost Niche Example (Online Dating): Websites like
"Gluten-Free Singles" or "Clown Dating" can be profitable despite
a tiny market share because the cost of replication is extremely
low
◆​ re-doing an existing website platform
➔​ Tactical Implementation: segmentation + 4P’s
◆​ Segmentation is used to tailor the marketing mix (Product, Price, Place,
Promotion) to maximize appeal and profitability for each segment
◆​ Pricing (Segmentation by Willingness to Pay)
●​ Disneyland Example: Price is segmented both geographically and
temporally.
○​ Goal: Practice price discrimination to capture consumer surplus
(charge each customer close to their maximum willingness to
pay).
○​ Segments:
 ◆​ Tourists/Out-of-State Visitors: Charged the highest price,
as they are often die-hards who require the experience
regardless of cost (low price sensitivity).
◆​ SoCal Residents: Offered substantial discounts (often tied
to specific days/restrictions) because they are more
sensitive to price and are flexible on when they visit, filling
park capacity on less popular days (Monday-Thursday).
◆​ Promotion/Communication (Same Product, Different Message)
●​ McDonald's Southwest Chicken Salad Example: The exact same product
was promoted using different themes based on the racial/ethnic segment
they were targeting.
○​ White Segment: Emphasized the salad as "fresh and gourmet."
○​ Black Segment: Emphasized the salad as "hearty and interesting,"
addressing a common insight that this group might not view a
salad as a satisfying meal replacement.
○​ Hispanic Segment: Emphasized "cultural flavors," highlighting
ingredients like cilantro and lime.
○​ Asian Segment: Emphasized "variety and newness."
◆​ Product (Designing Offerings for Specific Needs)
●​ Procter & Gamble (P&G) Laundry Detergents: P&G owns multiple
detergent brands to appeal to different psychological needs and price
points, filling the entire aisle.
○​ Tide: Appeals to high quality, long-term brand loyalty (often
passed down), and convenience (Tide Pods).
○​ Dreft: Appeals to parents of newborns ("Hopefuls") seeking
products that are perceived as organic, gentle, and un-chemical
(willingness to pay a premium for child-related safety).
●​ Coca-Cola (Diet Coke vs. Coke Zero): Two identical calorie-free products
targeting two different gender-based segments.
○​ Diet Coke (1980s): Developed when "Diet" was the mainstream
term associated with women's health and weight-consciousness.
Branding and taste were calibrated for this segment.
○​ Coke Zero (2000s): Developed to appeal to men who were now
also seeking calorie reduction but avoided Diet Coke due to its
historical association with women, its name, and its distinct taste.
Coke Zero used black/gray branding (masculine) and was
reformulated to taste more like regular Coke.
◆​ Place & Timing
●​ Tesla (Timing/Staggered Release): Releasing products sequentially from
high-end to mainstream.
○​ Strategy: Start with the high-priced Roadster and Model S
(2005-2012).
○​ Reasons: Captures high-margin profit from early adopter/die-hard
segment, which funds operational build-up, and establishes a
 premium brand image before releasing cheaper, mass-market
vehicles (e.g., Model 3).
●​ QuickVue Home Pregnancy Tests (Place/Product Differentiation):
○​ "The Hopefuls" Segment (Ready for Family):
◆​ Packaging: Features a baby, sold in smaller counts (25).
◆​ Place: Located near diapers, formula, and other
baby-related items.
◆​ Price: Higher per-test price (people will pay more for
products related to children).
○​ "The Fearfuls" Segment (Not Ready):
◆​ Packaging: Features only a woman, sold in bulk packs
(50).
◆​ Place: Located near birth control and medicinal/pharmacy
items.
◆​ Price: Lower per-test price.
●​ Hardee's vs. Carl's Jr. (Geographic): Same parent company, two brands
segmented regionally.
○​ Carl's Jr. (West Coast): Marketed as "Charbroiled Burgers,"
offering "restaurant quality in a fast-food setting."
○​ Hardee's (Midwest/South): Marketed as the "Charbroiled
Thickburger," appealing to a consumer preference for a "big, full,
proper meal."
➔​ The Consulting Segmentation Process (GBK Example)
◆​ common, structured approach used by consulting firms to deliver a segmentation
solution:
◆​ Get Up to Speed: Thoroughly understand the client's industry, competition, and
existing consumer knowledge.
◆​ Build Datasets: Survey consumers to collect data (especially psychographics)
and build the predictive segmentation model.
◆​ Cluster Analysis: Run statistical models (like k-means clustering) to divide the
market into distinct, homogenous groups.
◆​ Profile Segments: Analyze each resulting segment:
●​ Measure size and approximate market share.
●​ Use latent data (demographics, other behaviors) to create rich
descriptions of each segment for identification.
◆​ Targeting: Recommend to the client which specific segments they are best
positioned to serve and focus on.
◆​ Activation and Internal Embedding: Integrate the segmentation solution deeply
across the organization. This requires buy-in from departments beyond marketing
(e.g., R&D, Finance, PR/Advertising).
➔​ Activation: Naming, Imagery, and Internal Adoption (Mozilla/Firefox)
◆​ For a segment solution to be successful, it must be easily communicated and
adopted internally by all relevant teams.
 ●​ Creating Distinct Identity: Every segment is given a name, a color, and
imagery to make it sticky and easily understood across the organization.
●​ EX) The Mozilla/Firefox Example: Mozilla created a six-segment solution
to guide product development and marketing for its browser.
○​ Importance for Open Source: This is particularly critical for
open-source companies, where development teams are
decentralized. The segment profiles (e.g., "The Pragmatist," "The
Seeker") help coders and contributors understand exactly who
they are building features for.
○​ Result: The segments become internal dashboards, serving as a
constant reminder of the customer profiles, behaviors, and needs
they are trying to address.
➔​ Good segmentation
◆​ Relevant to firm’s objectives
◆​ Substantial: enough to profit
●​ Substantial: large enough to be worth the effort
●​ Low-cost cusinesses can compete in markets with may segments
◆​ Identifiable/Measurable: consumers can be identifies to segments
◆​ Accessible: must be able to reach them
◆​ Actionable: distinct segments have differential responses to firm’s marketing
actions
◆​ EX) good segment solution - Finance
●​ Data Used: Risk Score (X-axis) vs. Yield/Profit (Y-axis).
●​ Resulting Segments:
○​ The "Safe" Segment (Low Risk): They pay off debt perfectly,
resulting in negative yield for the firm (because the firm makes no
money on interest or fees).
○​ The "Profitable" Segment (Moderate Risk): They sometimes carry
a balance or incur fees, providing the highest positive yield for the
firm.
○​ The "Risky" Segment (High Risk): They often default or file
bankruptcy, resulting in negative yield for the firm.
●​ Actionability: The firm can now market different products to each group
○​ EX) high-interest products to the profitable segment, or basic
accounts to the safe segment
➔​ Segmentation criteria: Identifiable and accessible/actionable -> Look for a pattern:
Predictive of behavior
➔​ Beenfit segmentation: Predictive of behavior -> Look for pattern here: Identifiable and
accessible/actionable
➔​ Can be multiple segmentations within a firm
➔​ Segmentation witht he 4P’s:
◆​ Pricing: price discrimination, discounts
◆​ Promotions: ads, placement
◆​ Product: design, time fo release
 ◆​ Placement: geographic/retailer specific
◆​ Brand extension: product portfolio
➔​ Firm’s approach of consumer insights
◆​ Shore up foundational knowledge
●​ Ourside-In thinking
◆​ Build robust dataset for segmentation
●​ What are the needs/desires/behaviors
●​ Is it moral to market different products to different genders?
●​ Survey-based
◆​ Segmentation analysis
●​ How to cluster
●​ Segmentation bases
◆​ Sizing and profiling
●​ How big is each category + characteristics
●​ Segmentation and CLV drivers
◆​ Segment prioritization
●​ Which segment to focus on
○​ Key to segmentation is sacrifice
◆​ Activation
●​ How to bring segments to life + learn more about them
●​ More than nice to know
➔​ When tailoring products or marketing efforts to specific segments, marketers must tread
carefully to avoid crossing ethical lines.
◆​ EX) The Gray Area - Nerf: Differentiating products by color (e.g., blue vs. pink
Nerf blasters) might seem sexist,
●​ it could also be seen as appealing more directly to consumer preferences
and providing tailored joy
●​ the intent and consumer reception are key here.
◆​ EX) Crossing the Line - "Bic for Her": When segmentation becomes explicitly
patronizing or reinforces outdated stereotypes (like marketing pastel pens as
"BIC for Her"), it can lead to public backlash
●​ example of poorly executed segmentation that focuses on gender
stereotypes rather than actual differentiated needs for the product
function.
➔​ Cluster Analysis is the statistical tool used to group consumers into segments.
◆​ Unlike predictive modeling: where you have a known outcome, or 'Y' variable, like
"Did the customer buy?"
◆​ cluster analysis is unsupervised
●​ We do not know the "right" segment for a consumer
●​ algorithm must discover the optimal groupings using only the available
input variables (X variables).
●​ Goal: To create groups (clusters) where people within the cluster are very
similar to each other, and people between different clusters are very
dissimilar.
 ◆​ Types of Clustering
●​ Tree-Based Methods (Hierarchical):
○​ Agglomerative (AGNES): Starts with every person as their own
cluster and groups them upward (builds tree bottom-up)
○​ Divisive (DIANA): Starts with one large group and splits them
downward (builds tree top-down).
○​ The process visualizes the groupings as a tree (dendrogram)
○​ Same branch and lower the branch is the higher similarity
○​ Pick horizontal slice for cluster

●​ Distance-Based Methods (K-means): Groups are determined solely by

proximity to a central point (centroid).
○​ Diff colors for diff cluster represenartitives
○​ Cluster ppl on how close they are to their cluster representative
●​ Model-Based Methods (Density/Mixture Modeling): Assumes the data
comes from a mixture of underlying probability distributions (like multiple
hills on a 3D plot) and tries to mathematically identify the number and
shape of those distributions.
○​ Optimize weighted avg
○​ Bivariate formal distributions based on how many circles on graph
➔​ SPSS: point-and-click statistical software popular among market researchers from
roughly 1990 to 2010​
◆​ Rand 2 step cluster analysis
●​ 1) run kmenas w/ big K->many little consumer groups
●​ 2) each group as trees bottom and group tree uo
●​ Both and mixture model = increase stability
◆​ famous for developing a proprietary method called the "Two-Step Cluster
Algorithm."
●​ algorithm was a blend of methods, first using K-means (with a very large
K to find many small groups), and then using a Hierarchical (Tree-Based)
method to group those small K-means clusters upward into the final
segments
 ●​ This combination was believed to be more stable than using just one
method alone.
●​ NOT TYPE OF CLUSTERING​
○​ software package that contains various statistical tools, including
many different clustering algorithms
◆​ Clustering: It provides implementations of K-means,
Hierarchical (AGNES/DIANA), and other density-based
clustering methods
➔​ K-means Clustering Algorithm

◆​
◆​ K-means is the most common and visual distance-based clustering method,
requiring a predetermined number of clusters, K.
◆​ What it does: randomly assign ppl to cluster middle(centroid) -> assign ppl to
closest cluster middle -> move them to be in middle of everyone assigned
●​ Use nstarts: 10-20 for algorithm to start is diff places
●​ Some convert to best and some converge to local min -> give single best
one
◆​ Trying to find how far someone is from avg -> measures distance -> units matter:
scale all variables in dataset to have mean=0 and sd=1
◆​ The primary objective of the K-means algorithm is to minimize the Total
Within-Cluster Variation (W)
◆​ Total within-cluster variation: measure of the homogeneity (closeness) of the
points within all the identified clusters
●​ Hard Partition: K-means creates a hard partition, meaning every
consumer is fully and exclusively assigned to one cluster
(non-overlapping).
●​ Within-Cluster Variation (W): This is calculated as the sum of squared
distances between every point in a cluster and that cluster's central point
(the average). Minimizing W means making the points in each cluster as
close to their respective centers as possible (high homogeneity).
●​ Distance Metric: For continuous numeric variables, K-means typically
uses Euclidean Distance (the square root of the sum of squared
differences, like the hypotenuse of a right triangle).
◆​ The algorithm follows a simple, repetitive process:
●​ Initialization: Randomly assign K initial cluster center points (centroids)
within the data space.
●​ Assignment Step: Assign every single data point (consumer) to the cluster
whose centroid is closest to it.
 ●​ Update Step: Recalculate the position of each centroid by taking the
mean (average) of all the points newly assigned to that cluster.
●​ Repeat: Repeat the Assignment and Update steps until the cluster
assignments stop changing—the solution is said to have converged.
➔​ Considerations for K-means
◆​ The Local Minimum Problem: Because the algorithm starts randomly, it might
converge on a good solution that is not the absolute best possible solution (a
local minimum).
●​ Solution: nstarts: To increase the chance of finding the globally optimal
solution, we use the nstarts argument in R, telling the algorithm to run
the process multiple times (e.g., 10, 25, or 50) from different random
starting points. R then returns only the single best result (the one with the
lowest final W).
◆​ The Unit Problem (Variable Scaling): Variables measured in different units (e.g.,
age in years vs. income in thousands of dollars) would unfairly weight the
clustering process toward the variable with the largest numeric range.
●​ Solution: Scaling: All variables must be scaled before clustering to have
●​ a mean of zero and a standard deviation of one. This ensures all
variables contribute equally to the distance calculation.
○​ Scaled: all the same units
◆​ Cluster Labeling: The computer arbitrarily labels the final clusters (e.g., "Cluster
1," "Cluster 2").
●​ Solution: When communicating results, ignore the arbitrary label. Instead,
focus on describing the segment by its characteristics (e.g., "the largest
cluster," or "the segment with the highest income").
➔​ Determining the Optimal Number of Clusters (K) - how to find K
◆​ Scree plot->run kmeans 10 times w/ k=1 ->k=10 and plot clusters on horizontal
axis and plot within cluster sum(w)
●​ Adding clusters bg improvement and and some point small infelction
point(elbow)
◆​ Since the true number of segments is unknown, we need a method to estimate
the best value for K:
◆​ The Elbow Method (Scree Plot):
●​ steps to find the optimal K are:
○​ Generate the Scree Plot: You plot the number of clusters (K) on
the horizontal axis against the resulting Total Within-Cluster
Variation (W) on the vertical axis. The line should always
decrease.
◆​ Keep trying different values for K
○​ Identify the Inflection Point (The "Elbow"): You look for the point on
the graph where the line's rate of descent sharply changes—it
goes from falling steeply to flattening out significantly. This bend in
the graph is the "elbow."
 ○​ Select K: The value of K corresponding to the elbow is considered
the best choice.
◆​ Example below the K should be 3(right number of clusters)
●​ Bigger effects early on and smaller effects after
elbow point
◆​ Silhoutte scores: Mathematical formula way to evaluate cluster solution
●​ Right number of clusters? -> evaluate cluster solution
●​ HDB scan: new method
◆​ Greatest and Latest(2016) method: HDB scan

➔​ segmentiation: find structure of rows in dataset with cluster qanaysis and market
mapping
◆​ People who buy some product do so for diverse reasions
◆​ Firms response to main market feature, heterogeneity
➔​ Marketing pursue: match tru need and desires of consumers
◆​ Marketing plan: logical series of events to match
●​ Strategic plan: long terms/big pre plan -> find consumer needs, sowt
analysis
○​ Which consumers to focus on and create image of org to creat
●​ Tactical plan: what to do to put plan into action - micro
○​ Develop product in way to match segment
○​ Find right price based on comp
○​ Best way to distribute
○​ Communicate and promote is most appealing way
➔​ Segmentation: Tl, DR version - put people into groups
◆​ Within groups=similar
◆​ Across groups or b/w diff groups = diff
◆​ Important b/c if you try to cater to all = avg product
◆​ Either tailor product, the way we talk about it, promotuibs, pricing to appeal to diff
groups
➔​ How are groups heterogeneous​
◆​ Demographcus: age, race, income
●​ Easiest to collenct, wasn to segment just on this but
○​ Prince charles vs. ozzy
○​ EX) chris evans: 50’s, actor, famous, rich VS. ryan: more
entreprenaurial
 ◆​ Info and expeirences: abotu a product: underatnad pros of porcuct, if they
pruchased it before
◆​ Attitued: (+)/(-) feel from revuews/social media
●​ Better segmentation, but once segment based only on mind
○​ How do we find them iRL
○​ Cant survey every consumer
◆​ Psychogrpahucs: choices - lifestyle?relgious?art?
◆​ Needs: bigger car=family, streamer=bigger screen
◆​ Geographic: diff countries = diff norm = diff beaviorus
●​ Across states
●​ How to choose to store
●​ Diff area = diff wifi carriers
➔​ Good segementatuib
◆​ relevant: to firms goals
●​ Hwo to think about your customers
●​ EX) how to ad to diff consumers
●​ Can be narrow/vague
◆​ Substantial: segments can’t all be at individual level
●​ Need big enough groups to pofit and focus and care about
●​ Tiktok can b/ digital and have a lot of info on consumers at individual level
●​ Not practical at individual elvel
◆​ identify/measruable: can be assigned into segments
●​ How to find which segment someones in to carry diff marketing
◆​ Accessible: if main target elderly can’t do only online sales
◆​ Actionable: ppl respond different to diff marketing
●​ Need actioabilt thats didd across diff acting segments
➔​ Need to connect dots b/w easy to collect(demographuc) and what’s in ppl’s minds
◆​ Descriptive: start w/ demographic based segment
●​ Confirm these segments have psychographic/behavioral differences
●​ EX) young women, young men, old women, old men: good segmentation
if groups behave and respond differently based on prodcuts/industry
◆​ Benefit segmentation: get info on most predicit ve behaviro and tie back to
demographic differences o find these segments IRL
➔​ EX) walmart segmentation: can have multiple semgnets/firms
◆​ Income segmentation: cater to lower income
◆​ Price-sensitivity: can afford higher price for cars but comporsmise with cheaper
goods
◆​ Expand into area w/ less stores and become dominant - suburbs
➔​ How do firms use segments
◆​ Pricing: disneyland: $239 uf socal residement but $400 if nonsocal
●​ Geogrpahucal pricing change and everyday diff pricing
●​ Some willing to pay higher pruce to higher prices
●​ Some not willing to pay so have some days with lower prices(M-Thurs)
◆​ Promotions: mcdonalds salad promotions
 ●​ White: fresh and grommet
●​ Black: hearty and internesting
●​ Hispanic: highlight culturally relatable flavors
●​ asiansL variety and new
●​ Product: P&G laundry detergents - own many brands to fill up 1 aisle
○​ Tide: dorm friendly
○​ Dreft: newborns
○​ Maybe all rpdoduct have exact same content
●​ Prodcut release timing: tesla first cart roadstar(2005
->$100-$120k)->model S(2012->$80k)->model 3(2017->$60k)
○​ Build operations slowly to develop diff models
○​ Start w/ facienct sow when think tesla think fancy
○​ Start w/ highest pruce so most margin/car to build operations for
other cars
◆​ Release overtime: excited about tesla but only get
cheapest
●​ Release overtimeL excited about tesla then only 1
choice
●​ Placement: restaurants
○​ Wet coast: carl’s JR - charbroiled burgers
○​ East coast: same restaurant called Hardees: charibroiled
thickburgers
●​ Placement: Quidel preganancy
○​ Segment 1: hopeful for family
◆​ Have packaging with baby
◆​ Placed near other baby things
◆​ Higher pricing
○​ Segment 2: fearful for (+)
◆​ Package: no baby
◆​ Placed near meds/birth control
➔​ Helpind company segment
◆​ Get up to speed fast -> build datasets on which to segment on: survey
consumers’ needs/desires/behaviors -> segment analysis: seleclting variables to
do cluster analysis of to find segments -> sizing and profiling: how big and market
share/segment -> segment priortiization: client focu on few segments that they
well prioritize and well positions to address(targetting) -> activiation
➔​ Segmentation consideration:
◆​ Blue and pink: maybe meant like women=pink, but men still buy pink
◆​ Bic pastel for her: too far and bad
➔​ Clustering -> cluster analysisL consumers grouped to segments
◆​ Dont know right segment for consumer
◆​ Algorithm needs to find optimal groups w. Available variables(x)
◆​ Create clusters where within very similar and b/w cluster very different
◆​ Diff wats to evaluate clusters -> diff things being called best
Week 4:

➔​ Market mapping: use customer data to draw comp situation ad inform strategic
markeitng
➔​ Why market mapping
◆​ Undertnad brand/rpdocyt position in market
●​ EX) epsom salt
◆​ targetting/position
●​ Blue aporn: healthy eaters?convenvience?
◆​ Realt truth for customers is what they think truth is -> thei perception matters
◆​ Where to enter/waht products/feautres to develop
●​ Room in current market for differentiation/innovation
●​ MVMT watches
➔​ Mapping occurs overtime: b/c news and political events
➔​ When mapping do it using product attributes: what do consumer sperceive about
attributes
➔​ K-means clustering - Segmentation:
◆​ Goal: finding structure in the rows of dataset
◆​ Purpose: an unsupervised machine learning algorithm used for segmentation:
which people(rows) belong together
◆​ Output: groups of consumers - segments
➔​ Principal Component Analysis(PCA):
◆​ Goal: finding structure on the columns of the dataset
◆​ Purpose: primarily a dimension reduction algorithm: taking a lo of
variables(columns) and representing the info content in theme with less
variables(components)
◆​ Application: used on product attributes to understand how variables relate to
each other
➔​ Market analysis has to be based on consumer’s viewpoint
◆​ Decisions consumers make are based on the info in their head and what they do
with it
◆​ Regardless of what company/engineers believe to be true - customer’s
belief(perception) is the real truth that matters for marketing
◆​ EX) optical illusion of monsters
●​ 1 perceived as smaller than other
●​ Emphasized that what’s perceived(even if wrong) is most critical factor in
consumer decision-making
➔​ PCA commonly used to perform market mapping: process of visually representing
market based on attributed consumers care about
◆​ Goal: Understand product position
●​ Primary reason to map a market is to understand exactly where product
sits in competitive landscape
○​ EX) Dr. Teal’s Bath salths
 ◆​ Product: Epsom bath salts has an unclear position in the
store
◆​ Medicinal? Bath product? Spa/Relaxation product?
◆​ Conclusion: b/c product can be used differently by different
consumers - mapping helps determine the dominant
consumer view and use case
◆​ Goal: targeting and positioning decisions
●​ Mapping clarifies who the product is truly for and how it should be
positioned against comepittiorys
○​ EX) Blue apron
◆​ Blue apron delivers specific ingredients and recipe for a
meal
●​ Unlike instacart which deliver groceries/restaurant
delivery
◆​ Targeting ambiguity: who’s target consumer? Is it someone
who needs convenience?help/creativity in kitchen?healht
conscious people?
◆​ Positions: mapping helps decide which of these
benefits(convenvice/health/guidance) to emphasize to
attract the desire consumer base
◆​ Deciding where to enter/innovate - Gap finding
●​ Mapping identifies gaps in the market where a new or differentiated
product can succeed
○​ EX) Movement(MVMT) watches
◆​ Watch market had 2 extremes: cheap(plastic) and very
expensive(rolex)
◆​ Gap found: movement found a gap for watch that looked
nice(appeared to cost $1,000+) but was priced
affordable(around $200-$400)
◆​ Success: the company filled the space - growing into major
business by offering variety, fashion, and low cost over
mechanical perfection
○​ EX) BMW 3 series
◆​ Car market polarixed b/w non-luxury and luxury expensive
cars
◆​ Gap found: BMW created 3 series ad “introductory luxury”
●​ pulling new customers up into the luxury category
who could not afford the larger sedans
◆​ Impact: The product was wildly successful and has since
been copied by competitors (e.g., Audi A4, Lexus IS)
◆​ When mapping market, attributes that consumers use to evaluate products fall
into 2 categories
●​ Vertical Differentiation (Universal Agreement)
 ○​ Attributes where virtually all consumers agree that one direction is
better than the other
○​ Better to worse scale

Attribute Better Direction EX)

Price Lower is better All consumers prefer a lower

price, all else being equal.

Quality/Reliability Higher is better Cars requiring less service

(e.g., electric vehicle
drivetrains having fewer
components).

Speed Faster is better Faster CPU or processor for

computers.

Storage/Memory More is better Having more gigabytes or

terabytes of storage on a
phone or computer, even if
you don't use it all.

Comfort/Count Higher is better Softness of toilet paper or

thread count of sheets.

●​ Horizontal Differentiation (Preference/Disagreement)

○​ attributes where consumer preference varies widely, and there is
no universal agreement on which direction is "better." It's a matter
of taste or platform loyalty
○​ Your preferences vs my prefernences

Attribute Details EX)

Taste/Flavor Strong loyalty exists for Diet Coke vs. Diet Pepsi;
opposing options. people are often fiercely loyal
to one over the other.

Color/Aesthetic A matter of personal Some people like normal

preference. colors like blue; others prefer
something that stands out
 (e.g., the plain, customizable
aesthetic of the Cybertruck).

Platform/Operating System Driven by familiarity and Mac vs. Windows (or iPhone
switching costs. vs. Android). Users often
prefer the system they grew
up with, and switching incurs
a learning curve and "pain."

Game Console Loyalty based on ecosystem Sony vs. Nintendo (or the
or exclusive games. platform-specific debate for
gamers).

Location Choosing a place b/c its

closer rather than other
attributes

Complements A product is useful only if you old Lightning cable charger

have a compatible item for older iPhones, which is
useless for Android users

➔​ Market mapping is not a one-time activity because markets and consumer tastes evolve
◆​ Essential to map the market over time to track these changes.
◆​ EX) Ford Bronco
●​ The car was a competitor to the Jeep Wrangler and FJ Cruiser.
●​ After O.J. Simpson's famous police chase in 1994, the white Ford Bronco
received extensive, negative national press.
●​ Result: Sales fell, and Ford decided to discontinue it for 20-25 years due
to the negative association. Ford recently brought the Bronco back,
suggesting the market perception has finally shifted.
◆​ EX) Peloton and Sex and the City
●​ In the Sex and the City reboot (And Just Like That...), the character Mr.
Big dies of a heart attack shortly after riding a Peloton.
●​ Result: This was a high-profile example of bad press for the product,
negatively impacting perception among its core demographic.
➔​ Hotelling's Model of Differentiation (The Single-Dimension Street) -> horizontal
differentiation model
◆​ Developed by economist Harold Hotelling nearly 100 years ago to understand
product differentiation in a single dimension
◆​ dimension is often represented geographically (a "main street")
◆​ but it applies abstractly to any product dimension
●​ EX) "More Android" to "More Mac"
 ◆​ The Ice Cream Vendor Analogy (The Beach):
●​ Initial Logic: Two identical ice cream vendors (Gold and Blue) might
initially locate themselves far apart on a beach (a one-dimensional
space), believing they should avoid competition.
●​ Consumer Behavior: Consumers will choose the vendor closest to them,
creating a break point in the middle.
●​ Strategic Move: A vendor (Gold) realizes they can move toward the
competition (Blue) to capture a larger share of the consumers between
them while keeping all consumers on their far side.
●​ Equilibrium: Both vendors continue to move inward until they are located
very close to each other in the center of the dimension.
◆​ Implication: When products are identical (no horizontal differentiation other than
position/location),
●​ competitors will tend to cluster near the center of the market dimension to
maximize their market share
◆​ EX)
●​ Gas Stations: Shell and Chevron are often located patty-cornered or
directly next to each other, despite there not being enough demand to
support two stations independently, confirming the Hotelling model.
●​ Retail/Big Box Stores: In the Orlando market, Lowe's and Home Depot
stores are located within walking distance
○​ sharing parking lots or within one mile
●​ Airlines (Post-Deregulation): Before deregulation in the 1980s, when
pricing was controlled, competing flights (Delta, American) would depart
at the exact same time (e.g., 8 AM)
○​ After deregulation allowed price variation, they separated their
departure times (one earlier, one later) to cater to different
segments (business vs. vacation travelers)
○​ differentiating beyond just price.
●​ Political Spectrum: Candidates in a two-party system often do not
moderate to the center, as the most moderate candidate (C) would only
capture a small group of central voters,
○​ while the slightly more extreme candidates (B and D) capture
large segments of voters to their respective flanks.
●​ Vending Machines: A study found that if you place a Coke machine alone,
the consumer decision is "Do I want a Coke? Yes/No."
○​ If you place Coke and Pepsi machines next to each other, the
decision changes to "Do I want a Coke vs. a Pepsi?"
○​ works better for Coke, as it leads to more overall soft drink
purchases
➔​ PCA for market mapping
◆​ PCA as a Dimension Reduction Algorithm
 ●​ Purpose: Dimension Reduction. It takes a high-dimensional space (many
variables) and collapses the information into fewer, more meaningful
dimensions (Principal Components).
○​ New axes: a new set of coordinate axes that are mathematically
calculated to replace the original variables in your dataset
●​ New Axes are Ordered by Importance:
○​ PCA starts with N original variables (e.g., 9 product attributes) and
returns N new axes (9 Principal Components).
○​ The axes are ordered by the amount of variance (spread) they
capture in the data.
○​ PCA1: The most important axis, running in the direction where the
data is most spread out (captures the most variation).
○​ PCA2: The second most important axis, which is orthogonal
(90-degree angle) to PCA1
○​ J products(rows) and K attributes(rows) = J x K matrix
◆​ Math to find PC: calculate K x K matrix
○​ Diagonal: left to bottom right: vaiecnes of solo variables
◆​ How spread variable 1 is from its avg
○​ Off-diagonal:
◆​ (+): both vairables move together
◆​ (-): both go in oppositie
◆​ (0): o linear relation
◆​ Covariance of how both variables mvoe together
◆​ Understadning PCA mechanics
●​ focusing on Variation: If you start with 9 dimensions, you might only need
the first 2 or 3 axes (e.g., PCA1 and PCA2 -> can go in any direction(/|\-)
to summarize 80-90% of the interesting information
○​ The last few axes can be dropped because they capture very little
variation
●​ Technical Concept (Matrix Algebra):
○​ The algorithm calculates the Covariance Matrix of the product
attributes (K x K matrix).
○​ The Eigenvectors of that covariance matrix are the Principal
Components (the new axes).
●​ Interpretation is Key: PCA outputs these new axes without labels
○​ The analyst (you) must interpret the dimensions based on how the
original variables (attributes) map onto the new axes
◆​ EX) is PCA1 "Luxury vs. Economy," or "Speed vs.
Reliability"?
➔​ PCA in Practice: Academic and Marketing Examples
◆​ Teacher Quality Index:
●​ Input: 12 teachers rated on 9 measures of teaching quality (organization,
friendliness, fostering improvement, etc.).
 ●​ PCA Output: The 9 dimensions were collapsed into a single index of
teacher quality (PCA1).
●​ Result: This single dimension allowed the paper to rank teachers (e.g.,
"Teacher 9 is ranked best") without having to weigh the complex
trade-offs of the original 9 attributes.
◆​ Country Food Consumption:
●​ Input: Consumption data for different food types (fish, fruit, potatoes,
cereals) across four countries (England, Scotland, Wales, North Ireland).
High-dimensional data makes patterns hard to find.
●​ PCA Output: When reduced to a single dimension, the analysis clearly
shows that North Ireland is far apart from England, Scotland, and Wales
in its consumption patterns.
●​ Interpretation: Plotting the data on two dimensions (PCA1 and PCA2)
might reveal that PCA1 is "Nut Consumption vs. Meat Consumption"
(source of protein) and PCA2 is "Grains vs. Fish and Fruits"
(cultural/geographical factors).
●​ Result: Two components could explain 89% of the variation in food
consumption patterns across countries, suggesting that factors like
geography (landlocked vs. coastal) and culture are key drivers.
◆​ Car/Truck Market Mapping:
●​ Input: 9 different dimensions describing cars and trucks.
●​ PCA Output: The first two principal components (PCA1 and PCA2) are
used to make a market map (perceptual map).
●​ Goal: To visualize areas of high concentration (saturated competition) and
gaps (white space) in the market, which represent potential opportunities
for new product development or innovation.
➔​ Cons
◆​ How de we know its the 2 most important attributes
●​ What are the mos important, the make map on thesed dimensions
◆​ What about categorical attributes?
●​ So far only for continious
➔​ Selecting Important Attributes (MaxDiff Scaling)
◆​ main challenges in market mapping is determining which product attributes are
most important to consumers to use as the axes for mapping
●​ The Problem: Asking consumers to rank a long list of attributes (e.g., 10
or more) results in unreliable data because people struggle to hold that
many distinct preferences in their mind simultaneously
○​ The middle items get "jumbled up."
●​ Solution: MaxDiff Scaling (Best-Worst Scaling): This technique, often
used in consumer surveys, manages complexity while generating a
complete rank order.
○​ Task Design: Consumers are shown a small subset of attributes
(e.g., 3 or 4) from the complete list (e.g., 20 or 30).
 ○​ Choice: For that small subset, they must select the Most Preferred
(Best) and the Least Preferred (Worst) attribute.
○​ Design: The survey designer carefully balances which attributes
appear together and how often, ensuring every item is seen
multiple times.
○​ Scoring: The results are aggregated to create a score for each
attribute: (% Times Marked Best) - (% Times Marked Worst).
○​ Result: This MaxDiff score provides a highly reliable, rank-ordered
list of attribute importance (e.g., "handling and steering response"
vs. "blind spot detection"), which can then inform which attributes
should be emphasized in the product design or marketing
communication.
●​ 1. Collect attribtues to be tested
●​ 2. Make this choice task
○​ On survey show few choices and pick best and least fav -> fro few
choices get your preferences a little
◆​ Choices look random to consumers, but use design in
survey background: figures what attributes to show on
which page
◆​ Next: paired comparisons -> make sure options show up =
times
●​ 3. Collect data: fit into models and calculate
○​ (% times option was chosen as best/totoal times it was
shown)-((%times it was worse/totoal times show)
◆​ Handling Unknown or Hard-to-Measure Attributes (MDS)
●​ defining attributes of a product are vague (e.g., "quality" or "brand
feeling") or unknown, Multi-Dimensional Scaling (MDS) provides an
alternative to PCA
○​ Concept: MDS is a cousin of PCA that uses perceived
similarity/dissimilarity to reconstruct a map.
○​ Task: Instead of comparing attributes, consumers are simply
asked to rate how similar or different two products are (e.g., "On a
scale of 0 to 100, how similar are BMW and Acura?"). Netzer
◆​ Both gave 3 groups w/ some differences b/w cars put in
messages together and trade in behavior
○​ Process: The resulting matrix of similarity scores (the "distances"
between all products) is fed into the algorithm. MDS then works
backward, creating a spatial map where the distance between
points accurately reflects the stated similarity/dissimilarity.
○​ Result: This allows the creation of a perceptual map even when
the underlying attributes that drive consumer preference are
initially unknown.
➔​ Principles of dimension reduction and market mapping extend beyond survey data into
massive datasets derived from consumer behavior and language
◆​ Validating Market Structure with Multiple Data Sources
●​ Oded Netzer and co-authors performed a crucial study that validated the
structural integrity of market maps using two completely independent data
sources:
○​ Perceptual Data ([Link]): Researchers scraped over half
a million message threads from car forums. If two car models
(e.g., a Nissan Altima and a Honda Accord) were mentioned in the
same message, they were considered more similar in the
consumer's mind.
○​ Behavioral Data (J.D. Power/DMV): Researchers used trade-in
data to see which new cars were purchased when old cars were
traded in. Cars that are traded in for one another are also
considered similar.
●​ Finding: Despite the data coming from two totally different domains (what
people say vs. how people spend their money), the resulting market maps
showed the same basic structure—three clusters of vehicles:
○​ American-made cars (lower right)
○​ Premium European/Luxury cars (left)
○​ Predominantly Asian/Budget cars (top right)
●​ Conclusion: This cross-validation confirmed that the observed structure
(the grouping of brands) was not an artifact of one dataset but a robust
representation of the actual automotive market's competitive landscape.
◆​ The Rise of Embeddings: The idea of mapping relationships in high-dimensional
space is foundational to modern language models.
●​ Word Embeddings: This process is analogous to PCA for language. A
massive amount of text is processed, converting sentences into a sparse
matrix where words that appear together often are linked.
●​ Semantic Space: The resulting model creates a spatial map (the
"embedding space") where the distance and direction between words
hold meaning. For example, the vector that separates "man" from
"woman" is parallel and similar in magnitude to the vector that separates
"king" from "queen."
●​ If people liike at 2 prcuts in 2 session, then these 2 products thought to be
more similar than 2 products they didnt look at together
○​ Map all products in retailer - map figures out organic vs. inorganic
◆​ Jerky seen as more organic b/c higher protein thant
organic cookie
●​ Significance: This ability to capture semantic and relational meaning
through word position is a fundamental step toward building Large
Language Models (LLMs) that can understand and generate human-like
text.
◆​ E-Commerce and Substitution (Eccles)
 ●​ Using data from millions of shopping sessions, researchers mapped
product relationships based on the frequency of products being viewed
together.
○​ Insight: The resulting map reveals consumer substitution behavior
that transcends simple product categories.
●​ Example 1: Cookies: Organic cookies were often mapped closer to
non-organic cookies than to other organic health foods. This suggests
that for many consumers, the product category ("cookie") is a stronger
factor than the attribute ("organic").
●​ Example 2: Beef Jerky: Conversely, inorganic beef jerky was sometimes
mapped near organic health foods because, in the consumer's mind, it
shares the characteristic of being a natural, high-protein snack, making it
a viable substitute in that specific context.
●​ Application: These insights are valuable for store layout decisions (e.g.,
should orange juice be placed with other juices, or near breakfast foods
like milk and eggs?) and competitive strategy.

Week 5

➔​ Inverse demand curve: at each level diff people will buy

➔​ With P on y-axis and Q on x-axis have slope sloping down from left to right with E on top
having a line sloping and E’ prime another lien sloping down
◆​ Hard to find causal b/w P&Q b/c external factors: taxes, subs, comp, sots, ads,
politicies
●​ Causal b/w P and Q: steak example on changing price on menu
○​ Gives pure retaliation and trace out demand curve
➔​ Blue line: Inverse demand curve
◆​ Usually currency on x-axis and quantity on y, b/c flipped this is inverse
◆​ Company uses the curve by seeing how much they would pay to get the product
- willingness to buy
➔​ Demand curves: theoretical concept summarizing market response to price
 ◆​ Consumers maximize their utility function, by making choices according to their
preferences
●​ Each subject to income constraint and prices
●​ Prices are determined through equilibrium b/w firms and consumers
○​ Not set first
◆​ This causes individual demand for consumption set
◆​ Aggregate demand: sums all individual demand across consumers
◆​ Utility function properties = properties of aggregate demand fxn
◆​ Competitive, stable market w/ free entry (OR) Monopoly
➔​ Demand curves hard to estimate b/c may fluctuate with
◆​ Preferences, info, ads, costs, entry, comp, taxes
◆​ We want causal price-quantiity relation which requires exogenous price variation
●​ exogenous price variation: measures determined external to model, but
imposed by the model
➔​ Hard to see demand curves b/c companies dont keep moving around P to see how Q
responds
◆​ Uber knew exact demand of rides and supply of drivers
●​ Found equil of supply and demand
○​ Results: ppl who are seeing almost identical supply and medand
are facing different prices
◆​ Higher prices = lower acceptance rate
◆​ Ppl who saw identical supply and demand = lower ppl
purchasing
➔​ Can’t ask consumers what they would do if we changed price
◆​ Instead:
●​ Expert judgment
○​ Whomever regularly interacts w/ customers
◆​ EX) salesforce
○​ Very important in B2B
○​ A/B testing: change one thing in B group
◆​ If comp figures this out, they gain and learn cost structure
◆​ If comp finds your lowest price willing to go then can
launch temp price cut during experiment - skewed results
◆​ EX) ziprecuirter(due+misra): charges b/w $19-$399
●​ Results: profit ma at $250 and profit rose by 55%
●​ Price experiments
○​ May include in-makret tests, digital experiments, coupons, or
target price promotions
●​ Market research
○​ Conjoint experiments on surveys
○​ Make hypothetical choices on survey and fit stat model
○​ Can do things on survey that cant do IRL: gives purse cause and
effect b/w P and Q
○​ Most important marketing technique
 ●​ Demand modeling
○​ Statistical modeling on archival data
○​ Multinomial logit demand model
○​ Fit a stats model onto past choices consumers have made -> each
row = diff choice
○​ Pro:
◆​ Inexpensive, spending sometime plucjng right fata and fit
model
◆​ High confidentiality
◆​ Depends on IRL places where ppl truly spent money
○​ Cons
◆​ Need enough info -> info about you and your competitions
and what choices consumer choosing forms
◆​ Many companies do this but not much evidence of
modeling effects
○​ EX) how to motivate sales ppl to do better job: give incentives they
want
◆​ 9% increase of sales/month
○​ EX) Vegas to figure our to to get casino ppl to come back
◆​ Found best incentive and changed rewards system getting
$10-15M profit rise
➔​ Price experiments to discover demand
◆​ Can be easy(A/B testing: A in control and modify B a little to compare) or
semi-easy (Bandits: continuously allocated more traffic to better testing test
variation) to implement
◆​ This is observable to consumers: Consumers can see different prices at different
times or based on their history (personalized pricing)
●​ Change Purchase Timing: If a customer sees a low price now, they might
buy immediately, fearing the price will rise. If they see a high price, they
might wait for a sale
●​ Engage in Stockpiling: If the product is storable (like bulk goods or,
metaphorically, a long-term subscription), the consumer might buy more
than they immediately need when they see a low price
●​ Form Reference Price: Consumers learn what the "true" price of the
product is by seeing the range of prices offered. This "reference price"
then makes them less willing to pay a higher price in the future
◆​ Its observable to comp/suppliers/partners
●​ Interrupt Your Analysis: A competitor might launch a temporary,
aggressive price cut during your experiment to skew your results, making
it difficult for you to determine the true demand for your product.
●​ Learn Your Cost Structure: If a company sees you profitably selling a
similar product for a very low price, they can infer that your production or
procurement costs must be lower than theirs, or that you are willing to
accept a smaller margin.
 ●​ Game Your System: Partners (like ZipRecruiter’s clients) might
manipulate their behavior to trigger lower prices or otherwise exploit the
experimental parameters to benefit themselves.
◆​ EX) Dube&Misra(2023): Personalize pricing and Consumer welfare
●​ Partnered with ZipRecruiter
○​ Randomized firms to prices ranging from $19-$399
○​ Change base price from $99-$249
○​ Profits increase 55%
➔​ Conjoint: discovering demand
◆​ Is relatively cheap to survey consumers, and standard industry software handles
conjoint analysis
◆​ But is on survey/in a lab
●​ Stated behavoir can differ from actual behavior
●​ Elements of real world will be missing from experimental setting
●​ Design choices have to be made by researcher that may not mimic reality
◆​ May be the most important marketing technique developed in last 50 years
◆​ Used by almost all
➔​ Modeling on archival data: discovering demand
◆​ Relatively inexpensive
◆​ Confidential
◆​ Indices no cusotmer reactance or competitive response
◆​ Depends on real choices
◆​ Requires data, exogenous price variation, time, effort, training, trust
◆​ Assumed the future will resemble the past
◆​ Subject to untestable modeling assumptions
◆​ EX)
●​ Feldman 2021: field experiment on Alibaba shows revenue from sort orfer
by MNL outperformed ML model by 28%
●​ Nair 2017+Taylor&Zhang 2025: loyalty rewards with MGM casinos
resulted in $10-$15M incremental annual profit
○​ 20% improvement over RFM model
➔​ Multinomial logit model
◆​ Used to estimate demand
◆​ Linear regression: y is continuous
◆​ Logit/Probit: y in binary
◆​ MNL: y is categorical
➔​ MNL model rooted in random utility theory(RUT) where consumer chooses the product
that gives them highest utility
◆​ Utility: the utility a person(N) gets from a product(J) -> UNJ
 ●​ Vj: observed utility - systematic, measurable component that’s modeled
●​ Enj: unobserved error - random, unmeasurable component that captures
everything else influencing the choice(external)
◆​ Unobserved outcome: unlike linear regression, where Y variable is observed, the
actual utility number(U) is unobserved
●​ Model only observes final outcome of which product was chosen
➔​ Modeling the observed utility(Vj)
◆​ Systematic part of utility(Vj) modeled as linear function of product characteristics
●​ Linear form: Vj represented as linear combo of characteristics and their
weights(coeff):

●​ Product-specific variables(Xj): the observed variables(X’s) depend only on

the product(J) itself
○​ Don't depend on consumer N
●​ Price term(Pj): Price is pulled out specifically for emphasis due to its
importance in demand models, and it is assigned its own coefficient,
alpha
●​ Coeff(alpha and beta): These are vectors that represent the sensitivity or
"weight" of each input on utility.
○​ EX) large alpha means small change in price = large change in
utility
○​ Current Simplification: In this basic MNL model, all coefficients
(alpha and beta) are single numbers without consumer subscripts.
This means the model assumes a single, average market-wide
sensitivity to price and product attributes for all consumers.
◆​ Role of error term and probability
●​ The MNL model relies on a specific assumption about the distribution of
the error term to make the resulting purchase probability function
manageable
●​ Error assumptions: the errors (eNJ) are assumed to be additive and,
critically, independent across alternatives
●​ Gumbel distribution: the E terms are assumed to follow the Gumbel
distribution (also called the Type I Extreme Value distribution)
○​ Mathematical Convenience: This distribution is chosen purely
because it simplifies a complex, otherwise intractable integral
(involving J-1 multiplications and integrations) down to a simple
fraction
●​ The MNL Formula (Purchase Probability): The probability that
consumer N chooses product I(meaning Ui is max utility) simplifies to the
core MNL formula:
 ○​ Sensitivity: This formula shows that the probability of choosing
product I is sensitive to all inputs
◆​ I’s own characteristics, I’s price, and the characteristics
and prices of all competing products in the market (the
summation in the denominator)
➔​ Estimation and Demand Connection
◆​ Estimation Method: The parameters (alpha and beta) are estimated using the
Method of Maximum Likelihood (MMLE), which finds the coefficient values that
maximize the probability of observing the specific purchases recorded in the
real-world data.
◆​ Intuition: The computer tries to find alpha and beta values that make the model's
calculated purchase probability (Pi) very close to 1 for the products that were
actually bought and very close to 0 for the products that were not bought.
◆​ Demand Connection:
●​ Market Share: Since all consumers currently have the same sensitivities,
the market share for product I is simply equal to purchase probability Pi
●​ Quantity: Demand quantity is derived by multiplying the market share by
the total size of the market (e.g., 30% share 10 million phones sold). This
provides the final link between price (input) and quantity (output).
➔​ Major limitation: the IIA property
◆​ The independence assumption about the errors leads to the model's primary
flaw, which necessitates extending the model in later analyses.
●​ IIA Property: This is the Independence of Irrelevant Alternatives property.
It states that the ratio of the purchase probabilities between any two
existing products is unaffected by the introduction of a new alternative.
●​ The Red Bus, Blue Bus Problem (The Bug):
○​ Example: If a consumer chooses between a Car and a Blue Bus,
the ratio of Pcar to Pbluebus is 2:1
○​ IIA Violation: If a new Red Bus (a close substitute to the Blue Bus)
is introduced, the model maintains the 2:1 ratio b/w car and blue
bus
○​ Unrealistic Result: This forces the Red Bus to steal market share
equally from the Car and the Blue Bus
◆​ Realistically, the Red Bus should steal almost entirely from
the Blue Bus
 ◆​ MNL model fails this test because it treats all alternatives
as substitutes that are equally distant from each other
○​ Solution (Future Work): The solution is to extend the model (to
something like Mixed Logit) in order to allow consumer-specific
coefficients(alphaN and betaN) and correlated error terms
◆​ which will relax the IIA property and make the substitution
patterns more realistic
➔​ Limitation: price endogeneity
◆​ ver y important variable but didnt put into regression model
●​ Created OVB causing estimates to be wrong
●​ He x is important if correlated w/ included x and matters for outcome y
◆​ Im marketing price easy to measure w/ regression models VS quality hard so left
off the model
●​ But higher quality means higher price
●​ High pruce means less ppl buy more often
●​ Leads to OVB
➔​ For models we dont have R^2 telling u how well model fit to data - need other metrcus
◆​ Hit rate: % of times model predict correctly
◆​ pseudo-R^r: like R^2 w/ lowest=0 and highest =1
●​ 0.2-0.4: good fit model
●​ Higher number better

Week 6

➔​ Segmentation: finding structure in data rows using cluster analysis

➔​ Market mapping: finding structure in data columns using principal components
analysis(PCA)
➔​ MNL model: uses utility function: Unj = X’jB+apj+Enj

◆​ and specific error distribution(Type 1 extreme value) to simplify core choice

probability into tractable fraction:

➔​ Market share: sj=1/N summation n Pnj=Pj

➔​ A and b: weight senistivties
◆​ No subscript so avg $ sensitivity for full market is same
●​ An: diff alpha/person
➔​ Problem with basic MNL(Homogeneity)
◆​ Fixed Parameters: The model uses average price sensitivity (alpha) and average
feature weights (beta) that are constant for every person N in the sample.
◆​ Identical Probabilities: Since the parameters are fixed, the probability of buying a
product (Pnj) is the same for every person, which is not tremendously realistic.
◆​ IIA Violation: The model suffers from unrealistic substitution patterns known as
the Independence of Irrelevant Alternatives (IIA), also called the "Red Bus Blue
Bus" problem.
➔​ Heterogeneity
◆​ The Goal: To incorporate individual differences (heterogeneity) so that price
sensitivities (alpha) and feature weights (beta) can vary across consumers or
groups.
◆​ Why it Matters: Allows for targeted marketing actions (e.g., sending coupons only
to highly price-sensitive segments) and targeted innovations
●​ EX) creating different product sizes to address differing needs
◆​ Real-World Example (Transsion): The Chinese company Transsion achieved a
dominant 50% market share in Africa by recognizing a key consumer preference:
●​ consumers used the front-facing camera to take a selfie when testing
phones.
●​ Conclusion: Transsion adjusted the camera's default settings to be
particularly friendly to darker skin tones, demonstrating the power of
understanding a target segment's unique utility drivers.
➔​ Can’t just do +xn to model b;/c we donto observe individual utilities
◆​ Only observe chosen product
◆​ Amd both xn and xj would cancel out
➔​ 3 ways to implement heterogeneity
◆​ By segment(group-specific parameters)
●​ Approach: Divide the market into L segments (L=3) using cluster analysis.
●​ Mechanism: Allow the sensitivities to be segment-specific (alphaL and
betaL)
●​ Outcome: This is an improvement, as choice probabilities (Plj) will now
differ by segment, but they will still be the same for everyone within a
segment.
◆​ Interactions (Person-Specific Variables)
●​ Approach: Introduce person-specific variables (Wn = total minutes of
phone usage, age, income) and interact them with product-specific
variables (e.g., Price, Screen Size).
●​ Mathematical Necessity: Simply adding Wn to the utility function causes
the term to cancel out in the choice probability calculation. Therefore, it
must be introduced multiplicatively (as an interaction).
●​ Utility Specification: The utility function is structured to look like:
 ●​ Outcome: b/c Wn is unique to person N, the resulting choice
probability(Pnj) is person specific
➔​ Random Coefficients (Mixed Logit - Advanced)
◆​ Approach: Treat the sensitivities (alpha and beta) as random variables drawn
from a distribution (e.g., Normal, Lognormal).
◆​ Mechanism: The utility equation is specified with person-specific parameters
(alphan and betan).
◆​ Complexity: This is the most accurate and cutting-edge method (the focus of
Train's work), but it requires advanced hierarchical Bayesian statistics for
estimation and is considered too complex for the current class curriculum.
◆​ Benefit: It directly achieves the most realistic substitution patterns by explicitly
modeling the distribution of individual preferences.
◆​ The mathematical complexity arises when calculating market share (Pnj), which
requires an integral (an expected value) over the distribution of individual price
and feature sensitivities (alphan and betan) in the population, rather than a
simple average.
●​ Data Requirement: This approach requires longitudinal data—multiple
measurements (10-20 observations) for the same person over time.
○​ Rationale (Cereal Example): With only one observation, it's
impossible to tease apart whether a purchase was due to high
brand preference (beta for Cheerios) or high price sensitivity
($\alpha$ for cheaper cereals)
◆​ Observing multiple choices under different price scenarios
allows the model to learn the person's true underlying
sensitivities.
●​ Limitation for Class: The current dataset (phone purchases) is not set up
for this, as it contains only one purchase decision per person.
➔​ Conjoint analysis: data collection for demand models
◆​ specialized survey technique used to gather the necessary choice data to
estimate these advanced person-specific MNL models
●​ Significance: Cited as "the most significant development in marketing in
the last 40 or 50 years" by Professor Vithala Rao, due to its combination
of easy survey collection and powerful statistical analysis.
●​ Name Origin: "Conjoint" comes from the consumer evaluating product
features jointly when making a decision.
◆​ Conjoint methodology - sawtooth chart:
●​ Define Attributes and Levels:
○​ Researchers first determine key product attributes (e.g., Brand,
Price, Engine) and the specific levels for each attribute (e.g., Ford,
$35k, V6).
 ○​ The levels must cover competitors' current offerings and also
include potential, hypothetical options (e.g., a price point the
company hasn't yet used).
●​ Create Choice Tasks:
○​ Respondents are presented with multiple sets of hypothetical
products (called choice tasks) on a screen.
○​ These products are created from random (uncorrelated)
combinations of the attributes and levels.
●​ Collect choice data
○​ For each choice task, the respondent indicates the product they
prefer most or would purchase (often including a "None of these"
option).
●​ Conjoint output: Part-Worth utilities
○​ Once the choice data is collected, it is fit to the advanced MNL
model to estimate the sensitivities (or weights).
◆​ Present bar chart for 1 attribute w/ different bars
◆​ w/ 3 categorical variables -> turn into 3 dummy - in model
leave out 1 to not create perf multicollinearity: 1 variable
perf linear fxn of other so model cant tell unique
effects/variable = no unique coeff
◆​ Part-Worth Utilities: These are the estimated alpha's (price
sensitivity) and beta’s (feature weights).
◆​ Presentation:
●​ Utilities are often rescaled so that the average utility
across all levels within a single attribute sums to
zero (or averages to zero)
○​ This makes the utilities relative to the
average brand/level.
●​ Visualizing the utilities shows the relative
importance of different attributes.
○​ EX) In a phone plan example, data was
found to be the most important attribute,
followed by price).
○​ Google used it to figure out how their market share would be in
comparison to comp if they include feature vs. no
◆​ Once MNl model fit it gives you a and B sensitiivties
◆​ Benefits of conjoint analysis
●​ Causality: Because the hypothetical products are constructed using a
randomized, experimental design (attributes are uncorrelated), the
estimated sensitivities (alpha's and beta's) can be interpreted as the
causal effect of those features on consumer choice.
●​ Market Simulation: The fitted model allows marketers to set up any
hypothetical market scenario and predict the resulting market shares and
consumer choices (PNJs).
Feature Description

PRO: Causal Inference The experimental design (uncorrelated

attributes) allows the estimated sensitivities
(beta's) to be interpreted as the causal effect
of that feature on choice.

PRO: Simulation The fitted model can be used to simulate

market share for hypothetical products that
do not yet exist.

CON: Stated Preference Conjoint relies on stated preferences (survey

choices), not observed behavior (real
purchases with real money). This may not
perfectly reflect real-world actions since
people aren't spending real money.

CON: Assumption It assumes customers know how to make

consistent choices, even if they cannot
explicitly report their price sensitivity.

➔​ Conjoint Output and Product Innovation

◆​ The Minivan Example (New Product Development): Conjoint was used to
develop the minivan by combining attributes that didn't exist in the market before.
The key features consumers valued were:
●​ Large Sliding Door: Necessary for parents assisting children.
●​ High Number of Cup Holders: A highly valued, previously under-served
feature for families.
◆​ Archival Data vs. Conjoint: Conjoint is often faster and cheaper than analyzing
archival data because the survey is explicitly designed to collect all the necessary
information for model fitting, including non-chosen options.
◆​ The Power of Conjoint: This method is so powerful because it estimates a
demand model from economics but allows researchers to collect exactly the data
needed to fit the most advanced version of that model.
➔​ Strategic Decision Making (Google/Chris Chapman)
◆​ Google used Conjoint to predict market shares for a proposed new feature (e.g.,
AI integration in cloud services) versus competitor AWS.
◆​ Model Simulation: After fitting the MNL model, Google simulated four market
scenarios:
●​ Neither includes the feature.
●​ AWS only includes the feature (AWS gains significant market share).
●​ Google only includes the feature (Google gains market share, AWS drops
sharply).
 ●​ Both include the feature.
◆​ Result (Game Theory): The analysis showed that Google's predicted market
share was higher if they included the feature, regardless of whether AWS
included it or not. This revealed a dominant strategy for Google.
➔​ Political Economics (WashU/UCLA Economists)
◆​ Researchers used Conjoint to measure how the importance of political issues
(like abortion, immigration) affects people's voting decisions.
●​ Time-Series Analysis: They ran the survey at three points in time (before
the Roe v. Wade decision, after the leak, and after the final ruling) to track
changes in attribute importance.
●​ Finding: The importance weight (beta) for the abortion attribute decreased
significantly for Republicans after the ruling was favorable to them, but
remained high for Democrats.
➔​ Labor Economics (Labor vs. Leisure)
◆​ A Conjoint study measured the disutility (negative utility) of working different
numbers of hours.
●​ Finding: The decrease in happiness from working was linear up until
40–45 hours per week, but the disutility curve became steeper (happiness
declined faster) for every additional hour worked beyond that conventional
threshold.
➔​ Model Selection: The statistical technique of Cross-Validation (popular in machine
learning) will be used to help objectively select the "preferred model" from various
specifications.
➔​ Model Selection: The Overfitting Problem
◆​ When fitting multiple MNL models (e.g., should we add screen size or brand
intercepts?), a statistical tool is needed to objectively choose the "best" model.
●​ The Goal: To create a model that represents the big picture truth and
generalizes well, not one that is overfit to the specific quirks of the current
dataset.
●​ In-Sample Fit Limitations: Metrics like hit rates or pseudo-R-squared only
measure how well a model fits the data at hand. Adding more variables
always increases in-sample fit, potentially leading to overfitting.
●​ The Trade-Off (Bias vs. Variance):
○​ High Flexibility (Complex Model): Low bias (fits current data) but
high variance (performs poorly on new, unseen data, which is
overfitting).
○​ Low Flexibility (Simple Model): High bias (cannot fit the current
data well) but low variance (underfitting).
○​ The aim is to find the sweet spot that minimizes the total
Out-of-Sample Predictive Error (or Mean Squared Error).
➔​ Cross-Validation (K-Fold Procedure)
◆​ Cross-Validation is a model-agnostic procedure used to calculate out-of-sample
predictive performance (the Mean Squared Error).
●​ Better than hit rate and pseudo r b/c whose only give if good fit to data
 ○​ We want model that represents reality and can predict future
○​ Choosing best model: high flexible models= better fit to data
○​ Simpler models = better relationship capture
◆​ Bias variance tradeoff: need to use predictive and out of
sample stat = cross validation
●​ Lower flexibility higher bias=lower variance
○​ Prediction accuracy will be high error b/c high bias
○​ High flexibility = better fit but wont predict new data well
◆​ Will overfit current data and lower bias which increases
variance and higher erro
◆​ Bias: captures true relatio of data
●​ Higher bias = worse line of fit
◆​ K-Fold Cross-Validation Steps -> rectangles with 5 rectangles made of 5
squares, keep fitting data to model to 4/5 of the rectangles
●​ (actual-predicted)^2: out of sample mean squared error
●​ Split Data into K Folds: The full dataset is randomly divided into K equally
sized, non-overlapping subsets
○​ EX) K=5 folds, 20% each
●​ Iterative Training and Prediction: The procedure runs K times:
○​ In each iteration, the model is trained on K-1 folds
◆​ EX) 80% of the data
○​ The trained model is then used to predict the outcome(Y) for the
single held-out fold
◆​ The 20% of data it has never seen
●​ Calculate out-of-sample error
○​ After K iterations, every single row in the original dataset has an
out-of-sample prediction
○​ The mean squared error(MSE) is calculated using the difference
between the actual and predicted Y values
●​ Application
○​ By calculating the MSE for different models with varying flexibility
(e.g., 2 variables vs. 10 variables), you can objectively choose the
model that achieves the minimum out-of-sample prediction error

◆​ Effectively baalncing bias and variance

Week 7

➔​ Tucker: measruing customer reaction to price

 ◆​ Van west:ak 4 questions meant to figure fair price range
●​ For each question make curve
●​ T intersection you get the range
●​ Not the best method
●​ Not behavioral, good for new product, get wide range
◆​ Monadic questions: show some price 1 and others price 2 -> would you purchase
●​ Check %of yes
●​ Gives demand curve at different prices
●​ If more thant 2 point news more osumers because cant show many
options of price or they’ll catch on
◆​ Price elasticity: assume world operates at constant elasticity or measure elasticity
at 1 point
●​ Use it to find which $ to set
➔​ Pricing importance: avg US margin=8% which is each to change
◆​ Many say companies price too low
◆​ But may not charge higher b/c incentive
●​ Ppl whos et price wnt to keep their jobs by selling more
➔​ MNL model interpretation and fit: Product-Specific Variables for Exact Fit
◆​ The Rule: If you include a binary indicator variable (sometimes called an
"intercept") for every product in the model (leaving one out as a reference), the
model will match the observed product shares exactly.
◆​ Why it works: These product-specific coefficients (intercepts) capture all
differences between the products, including unobserved attributes, which allows
the model's predicted shares to align perfectly with the observed data.
◆​ Brand vs. Product Fit: If product shares are matched exactly, brand shares are
also matched exactly (since brand share is just the sum of its constituent product
shares).
◆​ Contrast (Brand-only Model): A model that only includes brand indicators (e.g.,
apple, samsung) can match brand shares exactly, but it cannot differentiate
between multiple products within that brand (e.g., Apple iPhone 1 vs. iPhone 2)
and will predict identical shares for them.
➔​ Understanding coeff interp. - endogeneity
◆​ Model without Price (Endogenous): In a simple model like choice ~ apple +
samsung, the estimated Apple coefficient is not the pure brand effect. It's
confounded by price (and other missing variables), a problem called endogeneity.
The coefficient represents the combined benefit of the Apple brand and the
penalty of its generally higher price.
◆​ Model with Price (Holding Constant): When price is added to the model (e.g.,
choice ~ apple + samsung + price), the interpretation changes to: "For the same
price, how much more do consumers prefer Apple over the reference brand
(Huawei)?"
◆​ Observed Effect: Adding price often causes the absolute value of the brand
coefficient (like apple) to increase (e.g., from 0.6 to 1.0) because the negative
 utility of the higher price is no longer embedded in the brand term; it's captured
by the separate price coefficient.
➔​ Reference level and Interp
◆​ Flexibility Constraint: You must omit one brand/product from the model (e.g.,
Huawei) because including all would lead to too much flexibility, making the
parameters unidentifiable (multicollinearity).
◆​ Relative Interpretation: The estimated coefficients (e.g., for Samsung) are always
interpreted relative to the omitted reference level.
●​ If Huawei is omitted, the Samsung coefficient is the preference for
Samsung relative to Huawei.
●​ If Apple is omitted, the Samsung coefficient would be the preference for
Samsung relative to Apple. The underlying models are equivalent, but the
estimated coefficient values change to reflect the new reference point
➔​ Traditional and advanced pricing strategies:
➔​ Importance of Price
◆​ Price is an easy lever to change (no need for product redesign or new
distribution).
◆​ A small 1% price increase can lead to a disproportionately large 11% profit
increase (based on an old, correlational McKinsey study), highlighting its
significant impact on the bottom line.
◆​ Managers are often hesitant to raise prices (and prefer lower prices for higher
sales figures) due to personal incentives (job security, appearing successful)
rather than true profit optimization.
➔​ Less Effective/Traditional Pricing Methods
◆​ Market-Based Pricing: Setting the price based only on the average market price,
often used when a product is treated as a commodity with little differentiation. It
ignores costs and specific consumer demand.
◆​ Competitor-Based Pricing: Setting the price based on a nearest competitor's
price, ignoring the firm's own product attributes and costs.
◆​ Cost-Plus Pricing: Setting the price by adding a fixed margin to the product's
cost. This is simple but completely ignores customer willingness to pay (demand)
and competitive conditions.
➔​ Van Westendorp Pricing Sensitivity Meter (Survey-Based)
◆​ This is a less-good, survey-based method that asks four specific questions to
triangulate a range of acceptable prices:​
●​ Price that is Too Expensive (Would not buy).
●​ Price that is Expensive (But might still buy).
●​ Price that is a Good Deal.
●​ Price that is Too Low (Question the quality; would not buy)
◆​ Limitation: It still relies on stated preference (what people say they would pay),
which is generally less reliable than observed behavior, and often results in an
unhelpfully wide range of acceptable prices.
➔​ Better Methods (Based on Behavior/Experimentation)
 ◆​ Monadic Questions (A/B Testing): Showing different consumers different single
prices on a survey and asking about their purchase intent. This provides
experimental data but is limited in how many price points can be tested without
needing a very large sample size.
◆​ Conjoint Analysis / Demand Estimation: Using sophisticated statistical models
(like MNL) based on observed choice data to estimate demand curves and price
elasticity, which is the most robust method for calculating an optimal,
profit-maximizing price.
➔​ Not great Supply-Side (Firm-Focused) Pricing Methods: traditional, older methods that
prioritize the firm's costs and profit goals over consumer demand
◆​ Cost-Plus Pricing (Most Common):
●​ Mechanism: The firm determines the cost to make the product (e.g.,
$1.00 to make a widget) and then adds a fixed target margin (e.g., 20%
margin) to set the price ($1.20).
●​ Pros: It's easy to do and ensures the company achieves a specific target
margin for profitability.
●​ Cons: It doesn't account for demand in any way. It completely ignores
what the customer is willing to pay or what competitors are charging,
potentially leading to lost profit opportunities or prices that are too high to
attract buyers.
●​ Real-World Example: The speaker's consulting company (GBK) largely
prices projects this way, combining estimated time costs with a tacked-on
margin that is also influenced by competitor bids.
◆​ Competitor-Based Pricing:
●​ Mechanism: The firm sets its price based solely on the price of its nearest
competitor.
●​ Critique: This method lacks differentiation; the firm just does what the
competitor does, ignoring its unique product value or costs.
◆​ Market-Based Pricing:
●​ Mechanism: The firm sets the price based on the overall market price for
the category, often used when a product is treated as a commodity.
➔​ A study of approximately 1,200 firms in Ireland (cited as 14 years old) revealed that
these traditional methods dominate price setting:
◆​ Cost-Based Pricing: Used by the vast majority of the firms (around 45% in the
table shown, focusing on the top panel).
◆​ Customer-Focused Pricing (Better Ways): Methods focused on demand or
value-based pricing fell into the "other" category, accounting for less than 10% of
firms at the time of the study, indicating a strong historical preference for simple,
firm-centric methods.
➔​ Demand-Side (Customer-Focused) Pricing Methods
◆​ These are considered "better ways" to set price because they start with the
customer and their willingness to pay, rather than internal costs.
◆​ Value-Based Pricing (EVC: Economic Value to the Customer):
●​ Goal: Set the price based on the customer's maximum willingness to pay.
 ●​ Framework:
○​ Identify the Next Best Alternative (Product Y) if the customer
doesn't buy your Product X.
○​ Calculate the Economic Value provided by Product X over Product
Y (this includes cost savings, quality benefits, and non-monetary
costs like convenience or time saved).
○​ The company then splits this total value/benefit with the consumer,
setting the price somewhere between the floor (cost to make) and
the ceiling (total EVC) to ensure the consumer still receives
consumer surplus (value remaining after the purchase price) to
incentivize the purchase.
●​ Use Case: Highly beneficial when a firm is low on data and needs to
make pricing decisions quickly (e.g., startups developing a business
plan), but often relies heavily on assumptions and guesswork to quantify
the "value."
●​ Example (The Batterizer): If a product (the Batterizer) saves the
consumer $4.00 worth of replacement batteries, the company should
price it above its own production cost ($1.00) but below the total $4.00+
benefit to share the value.
○​ Batterizer extends products life by 800% which is worth 8 batteries
and if each battery costs $0.5 -> product $0.l5(8) = $4
○​ Can be higher EVC than $4 b/c additional non prie benefits
○​ Can start with lower price to gain acceptance into market and then
increase prices
◆​ Experimental pricing
●​ Mechanism: Running controlled price experiments (like A/B tests in the
market or online).
●​ Benefit: Provides experimental variation that leads to causal effects (e.g.,
definitively showing how a price change affects sales/profit), such as the
case of ZipRecruiter moving its price from $100 to $250 based on
experimental learning.
◆​ Demand estimation
●​ Mechanism: Using statistical models (like the MNL model covered in the
course) to analyze actual observed choices by consumers, creating a
formal demand curve to find the profit-maximizing optimal price.
◆​ Customer Co-Determination:
●​ Mechanism: The customer or market determines the final price, seen
most commonly in bidding contracts (e.g., government highway
construction) or real-time auctions (e.g., ubiquitous online ad serving).
●​ Niche Area: Pay What You Want models (like Priceline), where the
consumer sets the price (often within certain constraints).
◆​ Conjoint method pricing:
●​ Give multiple options -> pick preference - for eahc attribute fet diff levels -
measure utltiles(a and B) that these attributes give you
 ●​ As long as price is product attribute 0 can use it to set high optimal $
●​ EX) attribute/level MNL coeff(B)(avg=0)
○​ Price
◆​ $70 - 18
◆​ $50 0.5
◆​ $30 17.5
○​ Data
◆​ 500 mb -25
◆​ 1 gp -10
◆​ 10 gb 12
◆​ Unlim 23
○​ 1. $/util: usually wit extreme $s
◆​ (70-30)/(17.5-(-18)) = $1.127/util
○​ 2. Pick target and reference and find utility change
◆​ Target: want phone with 1 gb
◆​ How to pruce it to a phone that has 500mb
◆​ 1 gb=-10 and 500mb = -25
◆​ Change in utility: (-25)-(-1) = 15 utils
○​ 3. Find value of change in utility
◆​ Change in utility($/util)
◆​ $1.127/util(15 utils) = $16.9
◆​ Gives EVC of moving from target to reference
◆​ Usually price it a little lower than thetis value
➔​Observed Behavior vs. Stated Preference
◆​ Observed Behavior (Preferred):
●​ Source: Uses actual choices consumers make in the market (e.g.,
demand estimation).
●​ Pros: Reveals true preferences, no gaming/lying by the consumer, and
preferred by economists.
●​ Cons: Can be difficult to acquire full market data (e.g., Samsung has data
on its own sales but not necessarily Apple's).
◆​ Stated Preference:
●​ Source: Asking consumers directly what they want or would pay (e.g.,
Van Westendorp, Conjoint, Monadic testing).
●​ Criteria for Reliability (Must be Met):
○​ Representative Sample: The consumers surveyed must accurately
reflect the target population.
○​ Knowledgeable Respondents: Consumers must be able to
realistically know the answer (they can't price a brand new,
unexperienced product).
○​ Honest Answering: The survey must be designed to mitigate the
"demand effect" (consumers giving the surveyor what they think
the surveyor wants to hear) and gaming.
➔​ Pricing factors: Price as a quality signal
 ◆​ Consumers often use the price itself to make inferences about the product's
quality, which complicates simple pricing models:
◆​ High Price = High Quality Signal:
●​ In situations where product quality is hard to judge (like sunglasses or an
unfamiliar brand of jewelry), a higher price can be perceived as an
indicator of higher quality, sometimes dramatically boosting sales (e.g.,
the pearl necklace example where doubling the price led to sales).
●​ Other Signals: Brand reputation, strong warranties (suggesting durability,
like Toyota's vs. Ford's), and costly signaling (like expensive Super Bowl
ads) all function similarly to imply quality or success.
◆​ Low Price / Promotions = Low Quality Signal:
●​ Price Promotions: Constantly running discounts signals to the consumer
that the product cannot be sold at its normal price and that its quality may
be suspect.
●​ Low Prices/Price Match Guarantees: Can suggest the product is a
generic commodity whose quality is not meant to stand on its own.
●​ Low-Quality Advertising: Can also signal a general lack of quality or
success for the product or dealership.
➔​ Consumer perception and signaling
◆​ Low-Quality Ads as a Signal: Low-quality advertisements (e.g., a local dealership
ad filmed in non-HD with a goofy floating inflatable) can signal that the retailer
lacks the resources or willingness to invest in high-quality advertising, which can
raise concerns about the overall business quality.
◆​ Costly Signals and Trust: Signals of quality are often costly to produce. If a
product seems "too good to be true" (great look, low price), and the quality
doesn't deliver, the signal backfires.
◆​ Consequences of a Backfire: A bad purchasing experience (e.g., the duffel bag
story) creates resentment, not just against the specific brand, but against the
entire purchase method (e.g., buying from social media ads), leading to loss of
future business across the platform/channel.
➔​ The Long Tail Effect: Online marketplaces introduce global competition and allow
retailers to cater to niche preferences outside of mainstream offerings (the "long tail").
◆​ EX) When pies were sold full-size, the most common choice (Apple) dominated.
When individual-size pies were introduced, people chose their preferred, more
niche flavors, and Apple dropped to third or fourth most popular
➔​ Non-Monetary Costs of Acquisition: The total cost of a product includes not just the
financial payment, but also physical effort (e.g., walking around a store) and mental
energy (e.g., evaluating a product's suitability).
➔​ Reducing Friction: Companies reduce these non-monetary costs to increase sales.
Amazon's "Buy Now" button (originally patented) significantly increased sales (estimated
$2.4 billion annually) by eliminating the friction of multi-step checkout (Add to Cart, View
Cart, Checkout).
➔​ Behavioral pricing effects($ tricks)
 ◆​ Left Digit Bias (Charm Pricing): Consumers perceive prices ending in $.99 as
significantly cheaper than the next dollar amount. People tend to focus on the
leftmost digit and round down, making $2.99 feel closer to $2 than $3.
●​ This effect causes a discontinuity in the demand curve (a sudden jump in
sales) right below every major dollar threshold (e.g., at $3.99 vs. $4.00, or
for Lyft rides).
◆​ Framing Discounts: The way a discount is presented heavily influences its
perceived size (innumeracy).
●​ EX) "30% off + an additional 40% off" (actual total discount: 58%) feels
like a much larger discount (70% off) than it is.
●​ EX) Consumers prefer "Buy one, get one free" over "50% off when you
buy two," even though they are mathematically the same.
●​ EX) Showing prices without commas (e.g., $1799 vs. $1,799) may make
the price feel shorter and thus lower, as the time it takes to read the
number is correlated with its perceived magnitude.
◆​ The decoy effect
●​ Decoy Effect Principle: Introducing a third, strategically priced option (the
decoy) that is clearly dominated by one of the original options causes
consumers to choose the dominating option more frequently, resolving
ambiguity between the original two.
●​ A clearly worse than B, dont know where B stand with C -> but rise in B
●​ Stabrucks EX) tall venti grande
◆​ Ambiguity Resolution: When faced with a difficult trade-off (e.g., $1.80 for 5/10
stars vs. $2.60 for 7/10 stars), the decoy (e.g., $1.60 for 4/10 stars) makes the
middle option (Brand B) clearly superior to the decoy (Brand A), driving sales to
Brand B.
◆​ Real-World Application:
●​ A study of a large online diamond retailer found that items that clearly
dominated a decoy diamond sold 1.8 to 3.2 times faster, increasing
overall profits by 14.3%.
●​ Williams Sonoma placed a low-quality bread basket on the shelf for $200
(the decoy) next to a high-quality one for $300 (the target). The $300
basket suddenly looked like a "phenomenal deal" by comparison, causing
its sales to soar.
➔​ Legal ways of price discrim
◆​ 2-part pricig: amazon: prime mebr free shipping+have to pay for solo products
◆​ Delivery time: earlier release products higher price
◆​ Reduce resentment: presnet as discount and never as paying premium
➔​ Contextual Willingness to Pay: A consumer's maximum willingness to pay changes
based on the environment of purchase, even if the product is consumed in the same
place.
◆​ EX) A person on the beach is willing to pay $5 for a beer bought at a rundown
grocery store but $10 for the same beer bought at a fancy resort, even though
the friend brings it back to the beach in both scenarios.
➔​ Price Obfuscation: The practice of hiding the true, final price until the consumer is highly
committed to the purchase, thereby reducing price sensitivity.
◆​ EX) Concert tickets and hotel bookings tack on fees (venue fee, resort fee) after
the customer has selected seats/dates and invested significant time, making
them unlikely to back out.
➔​ Price Disclosure Timing: Price information is shown at different times depending on the
context:
◆​ Before Decision: Stamps (highly price sensitive items).
◆​ After Decision: Michelin star restaurants or busy nightclubs (where the price is
secondary to the experience/convenience, and the cost appears only on the
receipt).
➔​ Preventing Resentment: When implementing price differences (e.g., to cover credit card
processing fees), frame the lower price as a discount for one group, never as a premium
for the other.
➔​ Price Discrimination Strategies: Legal and culturally accepted ways to charge different
customers different effective prices to extract more value:
◆​ Two-Part Pricing: A fixed fee for access plus a lower per-unit price (e.g., Costco
membership fee + low-margin goods; Amazon Prime subscription + free
shipping).
◆​ Timing of Purchase: Charging higher prices for earlier/exclusive access (e.g.,
movies in theaters first, then streaming, then free cable).
➔​ Framing Price Differences to Reduce Resentment:
◆​ Companies should always frame price differences as a "discount" for one group,
never as an "upcharge" or "premium" for another, to reduce consumer
resentment.
◆​ EX) Gas stations present a difference between cash and credit as a "cash
discount," even though it effectively covers the credit card processing fee.
➔​ Perceived Fairness: Consumers accept price differences when they are based on a
perceived system of fairness, such as loyalty or commitment.
◆​ EX) Coach passengers generally accept first-class upgrades for frequent flyers
(loyalty) because the system feels justified.
◆​ Contrast: Price differences based on irrelevant factors (e.g., charging different
prices based on a consumer's web browser—Chrome vs. Safari—as Amazon
once did) are not perceived as fair and lead to significant backlash.
➔​ The Bulk Buying Shift (Post-COVID):
◆​ For a long time, consumers were conditioned to assume that buying in bulk
yielded a lower price per unit.
◆​ Recent market changes (e.g., increased demand and stock-up behavior during
COVID-19) showed that, in some cases, the larger package became more
expensive per unit than the smaller package, defying the conditioned
expectation.
➔​ Avoiding Price Wars:
◆​ Smart firms differentiate their product and charge a markup based on that
differentiation.
 ◆​ Talking about competitors' prices only makes consumers aware of the competitor
and shifts the focus from product differentiation to price.
◆​ This leads to companies competing solely on price, which is good for the
consumer but bad for company margins and profit.
◆​ Firms should keep their costs top secret and avoid discussing competitor pricing.
➔​ Own Price Elasticity Definition (E): The measure of how much the quantity demanded
(Q) changes in response to a change in the product's own price (P).
◆​ Calculated as the percent change in Q divided by the percent change in P
●​ at a specific point P and Q

◆​ Interpreting Elasticity (Absolute Value):

●​ Inelastic Demand ( |E| < 1): Demand is inelastic when a 1% change in
price leads to less than a 1% change in quantity sold. Quantity doesn't
move much when the price changes.
●​ Elastic Demand ( |E| > 1): Demand is elastic when a 1% change in price
leads to more than a 1% change in quantity sold. Quantity moves
significantly when the price changes.
◆​ Economists' Preference: Economists favor elasticity because it is scale-free and
unit-free, allowing for comparisons across different products (e.g., gum vs. cars)
and different markets.
➔​ Price optimization models
◆​ The Constant Elasticity of Demand (Tucker Reading)
◆​ The Model: This is a mathematically convenient, specific functional form used in
economics: Q =

◆​ Log-Linearity: Taking the natural log (ln) of both sides transforms the equation
into a simple linear regression:
 ➔​
◆​ Elasticity as the Coefficient: In this specific functional form, the slope coefficient
(beta) of the log-linear regression is the price elasticity of demand.
◆​ Optimal Price Formula (Convenience): Given this constant elasticity (beta), the
profit-maximizing price can be easily calculated using the formula:

➔​
➔​Multinomial Logit (MNL) Model Optimization (Realistic Approach)
◆​ Deriving Quantity: In the MNL model, the market share for product J is calculated
first, which is a function of price. Quantity Sold (Qj) is then found by multiplying
the market share by the market size.

◆​ The Profit Function (pi): Profit is calculated as revenue minus cost (Quantity x
Price - Cost).

◆​ Finding the Optimal Price (Numerical Search):

●​ Unlike the constant elasticity model, the MNL model rarely results in a
simple function for optimization.
●​ The practical method is to test a wide range of candidate prices (P).
●​ For each candidate price, calculate the resulting profit (pi).
●​ The profit-maximizing price (P*) is the price that yields the highest
calculated profit.
◆​ Multi-Product Consideration: For sellers offering multiple products, the profit
maximization should sum the profits across all products (not just the product
whose price is changing) to account for cannibalization or complementary effects.

Week 8:

➔​ Brand purpose
◆​ Firms spend vast sums on brand building, and intangible brand capital often
makes up the majority of a company's sale value.
◆​ Branding helps build reputation (e.g., expecting high quality from a new Apple
product based on past experience).
 ◆​ It reduces search or transaction costs for the consumer.
◆​ It gives the firm monopoly power by lowering competition and capturing
consumer loyalty.
➔​ Celebrity alignment
◆​ Aligning with a celebrity can leverage their existing brand equity (e.g., Ryan
Reynolds buying a soccer team) to lift the company's brand.
◆​ The major risk is unfavorable association if the celebrity has a public scandal
(e.g., Kanye West or P. Diddy).
➔​ McGurk Effect: Demonstrates top-down processing, where visual cues (seeing the
mouth shape for "FA") override audio input (the actual sound is "BA"). We hear what we
expect to hear.
◆​ Top-down processing
◆​ Use existing knowledge to interpret sensory info -> brain sends down info to
sense: top-down
➔​ Color Perception (The Dress/Shadows): Our expectations about lighting and shade
manipulate our perception of color (seeing black/blue vs. white/gold).
➔​ Face Recognition: Humans are extremely good at recognizing faces, emphasizing the
importance of specific features like eyes and the mouth (demonstrated by the
upside-down Obama photo error).
➔​ Defining a brand
◆​ A brand is intangible—an idea or a collection of associations that lives in the
consumer's mind. It can be different for each person.
◆​ It is not just the logo, symbol, or physical product.
◆​ It is about relationships, not transactions.
➔​ Branding (The Process): The act of developing the brand and establishing a
psychological structure in the consumer's mind. The three big pillars of good branding
are Identity, Relevance, and Differentiation.
◆​ identiity: something it stands for
◆​ relevance: need to matter, connect, and build relation with people
◆​ Differentiation: dont call company a common word thatll be confused when
search up
➔​ Origin of Branding: Traced back to cattle ranchers who used hot irons (now liquid
nitrogen) to mark animals to prevent theft and allow buyers to easily identify animals
from the same, reliable ranch.
➔​ The psych of brand association:
➔​ Memory and Association (How Branding Works)​
◆​ Memory as an Auto-Associative Network: The core theory is that memory works
like a network where activating one idea (or node) cues other related ideas (e.g.,
trying to recall the happy birthday song requires internally singing the melody).
◆​ Retrieve things fro cues
◆​ Brand building: wen we activate 1 thing w/ brad so we want that one thing to
automatically activate other (+) things
◆​ Primacy effectL remember first few items
◆​ Recency effect: remember latest items
 ◆​ The Goal of Branding: To establish positive cues so that when consumers think of
related good ideas (e.g., speed, quality, celebration), their minds automatically
fire the link back to the brand itself.
◆​ The "Sleep" Test (Memory Experiment):
●​ The words shown were related to sleep (e.g., nighttime, slumber, drowsy,
dream, blanket).
●​ Many students incorrectly remembered seeing the word "sleep," even
though it was not on the list.
●​ Conclusion: The sheer volume of related concepts primed the idea of
"sleep" in the students' minds. Branding aims for this effect—to have
positive, cued associations lead directly to the brand.
◆​ The "Prefix" Test (Priming Experiment):
●​ Students were given the same prefix (OCT) twice but with different initial
prompts (SQUID and NOVEMBER).
●​ The typical result is the first prompt yields "Octopus" and the second
yields "October."
●​ Conclusion: A brand can use an associative cue (the "prompt") to
influence how a consumer interprets an ambiguous mark (the "prefix"),
demonstrating how context is everything in perception.
➔​ What can be branded
◆​ Answer: Everything can be branded because a brand is merely a collection of
associations in the mind.
◆​ Examples of Brands:
●​ Platform Brands: Airbnb (matchmaker, not product or consumer).
●​ Direct-to-Consumer Brands: Warby Parker, Kirkland (cutting out the
middleman to offer lower prices).
●​ Smart Brands: Google Nest, Amazon Alexa.
●​ Idea Brands: The Me Too Movement.
●​ Person Brands: Elon Musk.
◆​ The Patagonian Toothfish Example (Rebranding Success):
●​ Original Name: Patagonian Toothfish (unappetizing, "horrid-looking").
●​ Rebrand: Chilean Sea Bass (a non-fishy fish, making it versatile for chefs
to pair with any flavor profile).
●​ The new name is misleading ("sea bass" is a nonsensical term for a
non-freshwater fish), but it successfully changed perception and made it a
popular menu item.
➔​ Successful Branding Techniques​
◆​ Building Positive Associations:​
●​ Kodak (1995 Ad): Associated the brand with special, non-everyday life
moments (birthdays, weddings, graduation) to justify the cost and effort of
film photography.
●​ The Economist: Used a clever ad (a lightbulb turning on when walked
under) to convey that the magazine gives readers "bright ideas."
◆​ Avoiding Bad Associations:
 ●​ Twitter/X: After Elon Musk purchased the company and fired the content
moderation teams, the change in rules ("anything goes") caused major
advertisers to pull back.
●​ Advertiser Fear: Companies do not want their ads appearing next to "hate
speech" or "bad content" (e.g., a Disney ad next to a KKK discussion on
Reddit).
◆​ Logo Design (Hidden Meanings/White Space): Logos are small, recognizable
cues that embed meaning.
●​ Goodwill: The 'G' is designed to look like a happy face.
●​ Toblerone: The triangular chocolate shape represents the Swiss Alps; the
white space contains a bear (from Bern).
●​ VAIO: Combines an analog wave (VA) and the digital 1-0 (IO) to represent
the shift from analog to digital electronics.
●​ Baskin-Robbins: The number 31 (representing a different flavor for every
day of the month) is embedded in the 'B' and 'R' of the logo.
●​ Amazon: The arrow goes from A to Z (representing the wide variety of
products) and forms a smile (for customer happiness).
●​ FedEx: The arrow in the white space between the 'E' and 'x' suggests
speed and delivery.
➔​ Brand Recognition Test
◆​ Given different colors and figure out what matches
◆​ EX) brown = padres
➔​Creative Advertising (Enabled by Branding)
◆​ Once strong associations are established, a brand can use highly abstract and
creative advertising that doesn't need to explicitly name the product:
●​ Subway: An ad showing the Big Mac (bun in the middle) to suggest
Subway is a healthier alternative or competitor.
●​ KitKat: An image of a calendar full of online meetings, playing on their
slogan "Have a break, have a KitKat."
●​ Snickers: An ad using the familiar brown, white, and red packaging design
and colors, with the size resembling a Snickers bar on a taxi, without
showing the logo or name.
●​ Guinness: Stacked phones resembling a dark beer with white foam,
encouraging people to connect over a beer.
●​ Mercedes: Ran a campaign showing their three-pronged logo appearing
naturally in National Geographic-style photos, implying quality and
ubiquity.
➔​ Illusion of product differentiation: When products are fundamentally identical, their
success relies entirely on the associations built by the brand.
◆​ Vodka: Branding a "Neutral Spirit"
●​ Vodka is legally defined as a "neutral spirit" that is "without distinctive
character, aroma, taste, or color." The product is, by definition,
homogenous, forcing firms to create value solely through branding:
 ○​ Grey Goose (Luxury Brand): Focuses on exclusivity. Associations
include: French flag colors, frosted glass bottles, and the swan
logo, all conveying luxury. Advertisements push the slogan, "Drink
Luxury."
○​ Absolut (Fun/Social Brand): Focuses on local relevance and
socializing. Ads are bright, local (e.g., the LA swimming pool
shaped like the bottle), and emphasize a fun, accessible, and
social experience.
○​ Price: The perceived luxury (e.g., Grey Goose) is reflected in a
higher price premium compared to the more social or generic
brands
◆​ Bottled Water: The Ultimate Homogeneity
●​ The most extreme example is water, where consumers are asked to pay
up to 4,400 times more for chemical H₂O that they already have access
to.
●​ Success here is achieved by creating a perception of added benefits
(electrolytes, purity, natural spring sourcing) rather than delivering a
functionally different product.
➔​ How brands work: 4 core mechanisms
◆​ Brands are successful because they function in four distinct ways in the
consumer's mind:
◆​ Establishing Credibility: Consumers rely on brands as a signal of quality,
particularly when they lack expertise.
●​ The Price Premium for Credibility (Aspirin Study - headache meds):
Consumers will pay four times more for Bayer Aspirin (4 cents/pill) over
the generic Kirkland brand (1 cent/pill) for the same chemical compound
(acetylsalicylic acid). This is a price premium for the credibility and trust
associated with the original brand.
●​ Expert vs. General Consumer: Studies show that consumers who work in
the medical field (i.e., experts) are far more likely to buy the cheaper,
store-brand pain relief, demonstrating that their knowledge overrides the
brand's credibility. General consumers, lacking that knowledge, rely on
the brand.
●​ The Yogurt Taste Test Paradox: In a blind test, participants learned they
preferred the cheaper Roundy's yogurt over the expensive Chobani.
While they initially intended to switch to the cheaper brand, the data
showed that over the medium-to-long term, this behavioral change
dissipated, and they drifted back to buying the national Chobani brand.
The national brand's pull eventually outweighs the consumer's direct,
recent taste preference.
◆​ Rapid communication
●​ Brands use mascots, names, and visual language to instantly
communicate the product's function, target demographic, or benefit.
 ●​ Mr. Clean: Instantly communicates cleanliness, strength, and efficacy with
his appearance.
●​ Go-Gurt: The name itself is a blend of "Go" and "Yogurt," communicating
a product intended for "people on the go," packaged in a portable tube for
consumption by active children.
●​ Mountain Dew: Everything about the brand (extreme colors, font, and
advertising) conveys high energy and "extreme" caffeine intake, targeting
a specific, high-intensity consumer.
◆​ Personality and self-expression
●​ Brands develop distinct personalities that consumers use as a form of
self-expression.
●​ Mr. Peanuts: Was created to give the low-class, common bar snack a
"tuxedo and monocle" personality to distance the product from its humble
origins.
●​ Dove's "Real Beauty" Campaign: Successfully changed its brand
personality from traditional sex appeal to one of inclusivity by featuring
"normal-looking people," resonating deeply with its target audience.
●​ Insurance Industry (Homogenous Product, Distinct Personalities):
Insurance is inherently bland, making it a perfect canvas for personality
differentiation:
○​ GEICO: The Gecko emphasizes convenience and savings by not
having physical brick-and-mortar stores.
○​ Progressive (Flo): Focuses on being friendly, trustworthy, and
personable.
○​ Farmers (Professor Burke): Projects an image of wisdom,
experience, and reliability ("seen a thing or two").
○​ Allstate (Mayhem): Emphasizes the fear factor and the necessity
of protection against chaotic events.
●​ Brand Tattoos: The ultimate form of self-expression where consumers
permanently commit to brands like Nike or Harley-Davidson.
◆​ Gestalt (The "Extra" Factor): Gestalt means "the whole is greater than the sum of
its individual parts." In branding, this refers to the unexplainable, quantifiable,
psychological boost a preferred brand provides.
●​ The Coke-Pepsi fMRI Experiment:
○​ When participants drank soda without knowing the brand (blind
taste test), their brain activity was similar for both Coke and Pepsi.
○​ However, when participants drank the soda knowing which brand
it was, the one they preferred triggered a "bonus" of extra brain
activity (measured in areas related to reward and preference).
○​ This suggests that the brand knowledge (the collection of
associations, or the "gestalt") is separate from and adds
psychological value to the chemical product, leading to a
measurable difference in consumption experience.
➔​ Brand evaluation
 ◆​ Colors are Inalienable: Switching the colors of two competing brands (e.g.,
Coke's red/white and Pepsi's blue/white) makes both look immediately "wrong" or
"out of place."
◆​ Keller's Brand Report Card: A successful brand should score high on 10
fundamental measures, including proper positioning, consistency, and a sensible
portfolio.
◆​ Steve Jobs' Philosophy: The idea that a successful brand requires demanding
standards and a clear focus to set itself apart.
➔​ Gestalt: the principle that the "whole is greater than the sum of its parts," by presenting
further evidence that brand perception can chemically alter a consumer's experience,
even when the product is identical or swapped.
➔​Brand-Induced Preference
◆​ Peanut Butter Study: When researchers swapped the contents of peanut butter
jars, people consistently reported preferring the peanut butter that came from the
jar with the well-known brand name on it, even if that jar contained the inferior
product. The mere sight of the trusted brand name altered their perceived taste
experience.
◆​ Search Engine Study: Researchers created two search pages—one styled like
Google and one like Bing. When users searched, the back-end results were
secretly swapped (Google-styled page running Bing results, and vice versa).
Users consistently reported that the page that looked like Google offered better
results, showing that the brand's reputation for quality superseded the actual
functionality of the search engine.
➔​Brand Portfolio Strategies
◆​ When managing multiple products, a business must choose how to structure its
brands, typically falling into one of two categories:
●​ Branded House: The company uses a single, powerful master brand
across all its offerings. The brand equity is reinforced with every product.
○​ EX) Google (Gmail, Google Drive, Google Calendar) or Apple
(iPhone, Apple Watch, Apple Music).
●​ House of Brands: The parent company remains in the background, and
distinct brands are created for separate product categories. This is
common in consumer packaged goods (CPG) to target diverse
demographics without cross-contamination.
○​ EX) Procter & Gamble (P&G) owns Tide, Duracell, and Pringles,
with the P&G name rarely appearing on the product packaging
itself.
◆​ It's noted that the entire grocery store is dominated by a few major companies
(like Kraft Heinz) that collectively own vast numbers of individual brands.
➔​Brand Evolution and the "Debranding" Trend
➔​ Brands are not static; they must evolve to remain relevant. A recent trend observed
across major corporations is Debranding—the move toward simpler, flatter, and less
detailed logos.
 ➔​The Forces Driving Debranding
◆​ Mobile-First Design: The most immediate pressure is the need for logos to scale
effectively. A complex logo with gradients and shadows becomes illegible when
shrunk to a favicon or a tiny app icon. This drove companies like Burger King,
VW, Kia, and Pfizer to adopt simpler, flat (2D) designs.
◆​ Maturity: As companies grow from "playful innocence" to large, professional
corporations, their logos must mature to reflect the increased seriousness and
high stakes of their business (e.g., GoDaddy).
◆​ Versatility: A minimalist, neutral logo can act as a better "empty vessel" that is
flexible enough to span multiple product genres, allowing one symbol to cover 31
different flavors or diverse services.
➔​Case Studies in Rebranding Failures
◆​ Changing an established brand identity is extremely difficult, as strong brand
equity can lead to consumer rejection if the change is perceived as damaging or
confusing.
◆​ confusing.

Company/Syste Rebrand Attempt Outcome Why It Failed

m

HBO Multiple changes: HBO Executives eventually The name Max was
Go, HBO Now, HBO reversed course, often too generic and
Max, and ultimately returning to "HBO immediately
dropping "HBO" to just Max." sacrificed the HBO
"Max." brand's legacy of
premium, high-quality
content.

Gap Switched from its classic Rolled back within a Violated decades of
serif font in a blue box to week due to massive strong visual identity
a modern, flat logo in public outcry and that consumers relied
2010. hatred for the design. on.

UC System Replaced the traditional Abandoned within five It was criticized for
crest with a minimalist days. having no connection
blue and gold wave logo to heritage and
in 2012. visually resembling a
 "loading screen" or
"flushing toilet."

Google Confusing and Ongoing public Lack of clarity and

continuous rebranding confusion about which consistency; the
of communication apps app to use for chat, constant name
(Meet, Hangouts, Duo, video, or audio. changes created a
etc.) "total disaster" in
product identity.

Twitter Renamed to "X." Jury is still out, but it Forced adoption of a

sacrificed the globally name that lacked a
known verb "to tweet" cultural verb
and the cultural equity equivalent.
built around the
Twitter bird logo.

Cracker Barrel Attempted to remove the Brought the man back Consumers felt the
image of the old man after consumer guy was part of the
from the logo in backlash. brand's identity and
conjunction with heritage.
modernizing the in-store
atmosphere.

Week 9:

●​ Research question: how well does each factor explain merchant revenue
●​ The revenue that comes into the company(merchant revenue) comes from number of
stores that the company operats*the number of cusotemrs that come/store*avg number
of transactions/customer*for each translation the avg amount of money they spend
○​ D = S * C/S * T/C * D/T
○​ Run regression on 4 different acts and the beta coe3ffieicnt from the regression
will tell you importance/fctors in explaining revenue of merchant
○​ Run multiple regressions -> got a matrix
○​ Fixed effect: if put gender into model it would come out as 1 binary viarble for
each category
■​ But need to leavr out one variable for multicollinearity
■​ EX) what year are you: b/c 4 years, code in 3 dummy variables and put in
3 binary vairuables into regression
 ■​ Each coeff you get on that variable is something that shifts the intercept
for that group of people
■​ Fitting the model and putting some variables in fixed effects
○​ Findings:
■​ Number of customers explains most vairiaiton in merchant revenue
■​ Then size of trqsaction more important that number of trqnsactions made
●​ If more expensive price items then size of transaction hold more
weight thant buying from 7/11
●​ Market size: how big is the market that we operate in
○​ M: total size of the market - number of people who have some core needs that
you may serve which can be measured in dollars/units/volune
■​ EX) oil and gas industry dont measure by number of people they serve
but by the thousands of barrels of oils they can produce
○​ Trying to estimate the market size:
○​ Top-down: look at total market and try to estimate each person that has this need
■​ Would it be possible to serve all of these people?
○​ Bottom-up: some people are currently being served. They have a need and they
are being served by some company
■​ How much more room for groweht is there above the service that’s
already happening?
○​ Ask expert: give us an estimate of the market size
○​ None of these are accurate on their own so need to take average
●​ EX) Total addressable market is everyone who needs a amttress ~ same size as the
served market(everyone at this point has a market)
○​ Supposed to replace mattress every 7 years - but we may do less often
○​ Huge range for mattress costs: we take average of $300
○​ Assume population of 340 million
○​ CANT just do 340 milion/7 B/C NOT every single person needs a mattress
■​ Creates overcounting b/c couples amy sleep together
■​ Temporary need for mattresses: hotels/dorms
■​ Extra beds, maybe each eprosn has multiple houses so more
mattresses/person
■​ Assume the adjustments all offset each other so every 7 years mattress is
getting replaces and there’s 1 mattress/person of the 240 million people
●​ 340 million/7 years = 48.6 million mattresses
■​ Costs $300/matrress so
●​ $300*48.6 = $15 billion industry is the size of the market
●​ Charge different pricing to try to get as much consumer surplus as possible b/c everyone
is willing to pay a diff price
●​ EX) Google trends
○​ The graph goes up and down constnalty - the ups are peaking during college
szn, the low times are when no one is applying
●​ Market size can change overtime
○​ mass adoption - lack of adoption - start adoption again
 ○​ Can be caused by political events/demand side effects where consumer out of
work so purchase less, supply side with mandates or making prices increase
○​ Rate of adoption for newer things tend to be faster, even faster with software
products
●​ Diffusion - spread of adoption: Frank Bass - borrows idea from the spred of
disease(epidemiology) - idea tht spread of technology into a group of consumers can be
modeled like the spread of disease among group of population
○​ Starts idea of innovation and diffusion of products within market
●​ Bas model equation Inputs:
■​ t: time period
■​ M: market size - number of adopters
■​ N(t): number of new adopters in period t
■​ A(t): number of accumulated customers who adopted before period t
●​ Cumulative sum of all people who have adopted so far
●​ At the start of the period - everyone who bought up until period T
so doesnt include N(t)for same peiro
■​ R(t): number of remaining customers who can potentially adopt in period t
●​ Total market size-A(t)
○​ Speed at which people adopt
■​ p: coeff of innovation - curios about product so they are going to buy it
■​ q: coeff of imitation - as some people adopt, how that affects other
adopters
●​ Try to control network effect: idea that if some people have this,
then its going to make others want it to

■​
○​ 1st equation: relationship b/w size of market M and breaking it down to total
nu,ber of people that have accumulated and those who need to accumulate
○​ 2nd formula: rate of adoption of people and the P and Q seen as rates
■​ The total number of people adopting in period T over those tha are left to
adopt that haven’t adopted yet
■​ Percent(p): the number of people who will buy the tech the next time
period out of those that are remaining to purchase
○​ Need to specify values like M, P, Q
■​ EX) total market size - 100M users, P=2% of adoption, Q=0.4, simulate
BAS over 25 time periods(J)
●​ 1st period no one has adopted yet
○​ Cumulative number of adopters before the first time
period(A)=0
 ○​ Use the equations: A(1)=0, R(1)=M-0=M, N(1)=Mp
■​ Use second equation for N
■​ Cumulative number of people who adopted=0=p ->
N(t)/R(t) = 0+q(A(t)/M) = N(1)=Mp
■​ Next create vectors for A,R,N and each length of 25
to store the values
○​ Code follows same math(Bass mode simulation slide)
○​ During the next time period: total people who accumulated
in period T is the number who accumulated until the last
time period+in the last time period
■​ T=5 -> want everyone who is accumulated up
through the 4th time period+people who
accumulated during 4th time period = everyone
who hs accumulated until start of fifth time period
●​ Take the equations and write them into code->run the for loop and
get values of A,R,N for each of the first 25 time periods
●​ Chart on the slide: plot 1-25 on horizontal axis and number of
people who adopt in each period on vertical xis
○​ Interp: some scorpion and rhe number of adopters/time
period will grow
○​ Added to natural rate of adoption is something tha depend
on total number of people who have adopted so-called
○​ As tech starts to diffuse into market, more people jump on
board, as we start to saturate the market there’s fewer
people left who cna adopt the product
○​ Rate of adoption has to come down because there’s fewer
people left to adopt so upside down curve
■​ If plotted total cumulative number of adopters(instead of per period) =
cumulative sum(bottom chart on the slide) -> slower start and then picks
up faster, but as we start to fill up the market its going to tail off
●​ Can get different shapes by playing with diff numbers for P and Q
○​ Using BASS IRL: collect data on N(number of people who have adopted) and we
can use the N’s to make A’s = cumulative number of people who have dopted
■​ How many people have bought the product in eahc time period: we can
estimate P,Q,M - once we estimate, we can project forward
■​ Each of the adopter/time eperiod hasa an A, B*A, (C*A)^2
■​ From running regression we will know A,B, and C
■​ N: outcome variable
■​ Adopters: first period 3 adopted, 2nd period 6 adopted, so start o 3rd
period: 6+3=9 adopters
○​ If smeone collected the data, over some periods of time - how many people
bought the product?
■​ Take this vector od data and make the cumulative nu,ber of people who
bought the product
 ●​ Bass model algebra: take the 2 bass model equation and put them together to rewrite
the relationship
●​ If dont want to leave it upt ot model to make guesses about market size [ find P and Q
from BASS model, if you know M
●​ BASS model rewritten as

●​
●​ A little but of error b/w left and right side fro any values of P and Q
●​ Use nonlinea least squares: take left side of equation, N - the right side
○​ For eahc time period look at the difference, square it , and sum differences = sum
of squared errors from models
○​ Pick P and Q to minimuze sum of squared errors
an Yavorsky: If you had then done that, so you collect a little bit of data, like we put on the board,
use that data to fit the BAS model, bind P, Q, and M,

Dan Yavorsky: Once you've collected that data, you could then project forward to see what the
future is going to look like. So we estimate, we collect some data, we estimate the BAS model
parameters.

Dan Yavorsky: And then we're going to project forward, the dots are red on the projection, what
the… what the growth is going to look like in the industry, according to P and Q, and where growth
is going to taper off, and the market's going to be saturated, like, how big is the market?

Dan Yavorsky: And we're going to learn all this, P, Q, and M, 3 very interesting things to know about
the market, just from the little bits of adoption that we see at the beginning, and the structure that's
built into the VAST model.

Dan Yavorsky: So, when Frank Bass went to go do this in his paper, this was, like, he wrote the
paper in the 1960s, and the data was from, like, the 1940s, 50s? Yeah, 1947 to 1961.

Dan Yavorsky: So, old data, but…

Dan Yavorsky: After he does this, he then projects forward, so he has data before what's plotted,
and that's the… what he's fit his model to. And then he predicts forward what his model thinks is
going to happen in the industry, and then he charts out what actually happened in the solid line.

Dan Yavorsky: And you'll see that his predictions aren't perfect, right? His model a little bit
over-predicted at first, and then a little bit under-predicted in the middle.

Dan Yavorsky: But by and large, his model is fitting these adoption curves really pretty well.

Dan Yavorsky: This is the same reason the multinomial logic became so popular. I mean, first of all,
it's a good model, but second, its grand debut on the BART system in San Francisco was really
accurate. It did a really good job. It proved itself as being useful. And Frank's model does the same
thing.
Dan Yavorsky: Right? He's showing you that this does actually a very good job at predicting,

Dan Yavorsky: the future of the number of sales that are going to happen in each time period, and
also giving you, like, the… when the market becomes saturated and sales will start to decline. Kind
of shows you this rate of adoption in the industry.

Dan Yavorsky: Okay, if you're trying to put a C and Q in M,

Dan Yavorsky: The statistics are actually not very hard. You run one linear regression model, like
you learned in ECOM 120D, and you solve the three equations for the three unknowns, and you
get P, Q, and F.

Dan Yavorsky: If, however, you don't want to leave it up to the model and the little bit of data you've
collected to make an important guess about the size of the market app.

Dan Yavorsky: you might have, like we did with the Mattress Global Industries Trend Report, some
other very good guess at M that you get from some other data.

Dan Yavorsky: If you have that, if you know M, so to speak, then all you need to do is find the P
and the Q from the BAS model.

Dan Yavorsky: This feels like it should be easier. We have fewer things to learn. Before, we were
learning PQM, now I just want to learn P and Q. It seems like it should be an easier question, but in
fact, it needs slightly harder, statistics.

Dan Yavorsky: And it's a chance for me in the code to show you something that's,

Dan Yavorsky: I think it's pretty cool, so I'm going to show you on the slide, and then we'll get to
doing the code together.

Dan Yavorsky: Which one do we want to use in the code?

Dan Yavorsky: Hmm… what to do here?

Dan Yavorsky: The Bass model equation can be written as this N is equal to P plus QA over M
times R. So we've seen this relationship before.

Dan Yavorsky: We've collected data on N, and we think it should fit these things on the right-hand
side exactly.

Dan Yavorsky: But we fixed M. We know M now, so to speak.

Dan Yavorsky: So what we have…

Dan Yavorsky: For any values of P and Q, we have something on the right-hand side we can
calculate that may or may not exactly equal the left-hand side.

Dan Yavorsky: There's probably going to be a little bit of an error between the left-hand side and
the right-hand side.

Dan Yavorsky: So what we're going to use is something called, nonlinear least squares.
 Dan Yavorsky: The same idea as least squares. We're going to take the left-hand side of the
equation, N, minus the right-hand side, whatever we calculate involving M,

Dan Yavorsky: It may not exactly equal, so there's going to be a little bit of error between the
left-hand side and the right-hand side.

Dan Yavorsky: And we're gonna… for each time period, we're gonna look at this difference, we're
gonna square it up, and we're gonna take the sum of those differences. This is the sum of squared
errors from the model. We want to pick P and Q in order to minimize the sum of squared errors.

Dan Yavorsky: So I'll show you in code how to do some nonlinear least squares as the last thing
today.

Dan Yavorsky: Okay, Bass's model doesn't have a lot to say for why adoption happens the way it
does, right? He borrowed this simple mathematical model from the spread of disease, and he
applied it to the rate of adoption of a technology in marketing.

●​
●​ Market Sizing and Industry Analysis
●​ Target Audience for Research Reports (B2B):
○​ Expensive industry reports (e.g., mattress trend reports priced $3,000 - $7,000)
are sold to businesses, not curious consumers.
○​ Strategic Buyers: Companies (like Casper, Purple) need detailed industry data
to understand their market and competition.
○​ Investment Buyers: Investors managing large funds (e.g., $500 million) require
reports to allocate capital sensibly across industries.
●​ Pricing Strategy and Consumer Surplus:
○​ Companies often use "Good, Better, Best" (3-tiered) pricing (e.g., single-user,
multi-user, enterprise license) to practice price discrimination.
○​ The goal is to extract maximum consumer surplus by offering different
products (access/licensing) tailored to different customers (e.g., an analyst vs. a
large corporation's research team).
○​ The middle option often serves as a focal point, sometimes alongside a decoy
option (like an overly expensive bottle of wine) to push consumers toward the
next best-priced item.

📈 Analyzing Market Trends Over Time

Google Trends Data
●​ Purpose: To gauge relative interest (search volume) over time and compare multiple
search terms.
●​ Data Scaling: The data is indexed, not absolute. The peak search volume in the period
is set to 100, and all other points are scaled as a percentage of that peak. The exact
number of searches is not revealed.
 ●​ Long-Term Trends: Data over very long periods often show a slight downward trend
for a specific term because the total number of searches on Google is constantly
increasing.
●​ Real-World Example (UCSD): Exhibits predictable annual spikes (during application
season) and smaller cyclical spikes (start of academic quarters).
●​ The "COVID" Effect: Any time-series data covering 2020-2021 will show anomalies,
which must be considered (e.g., as a factor) in statistical modeling.

Technology Adoption Curves (S-Curves)

●​ Shape: Adoption often follows an S-curve: slow initial uptake, rapid acceleration, and
then saturation/tailing off as the market matures.
●​ Adoption Speed: Newer technologies (post-color TV, cell phones, internet) are
adopted significantly faster and more steeply than older ones (e.g., telephone,
electricity).
●​ Accelerating Factors:
○​ Observability: The ability for consumers to see others using and benefiting from
a product (accelerated by media like color television) drives imitation and faster
diffusion.
○​ Replacement: New technologies often displace existing, inferior products (e.g.,
cell phones replacing landlines).
●​ Historical Pauses: Adoption curves for technologies in the 1930s and 1940s often show
a flat area or dip due to the Great Depression (demand-side effect) and WWII
(supply-side effect, as resources like steel were diverted to war efforts).
●​ Software Diffusion: The adoption rate for software (e.g., ChatGPT reaching 100 million
users in two months) is dramatically faster than for physical goods.

🌊 Bass Diffusion Model (BAS)

The BAS model describes the adoption of a new product by modeling its spread as similar to
the spread of a disease (epidemiology).

Key Components
●​ $M$: Total market size (maximum number of potential adopters).
●​ $N_t$: New adopters (first-time purchases) in time period $t$.
●​ $A_t$: Cumulative adopters at the start of period $t$.
●​ $R_t$: Remaining non-adopters ($M - A_t$).
●​ $P$ (Coefficient of Innovation): Represents adoption driven by external factors (e.g.,
media, curiosity, core need). A natural rate of trial.
●​ $Q$ (Coefficient of Imitation): Represents adoption driven by network effects or
word-of-mouth (seeing others use the product).

Core BAS Equation

The equation relates the proportion of remaining non-adopters who will adopt next period to the
two coefficients and the current adoption level:

$$\frac{N_t}{R_t} = P + \frac{Q}{M} A_t$$

Fitting BASS to Find $P$, $Q$, and $M$ (Standard Method)

1.​ Data Preparation: Collect data on $N_t$ (new sales) over time. Calculate $A_t$
(cumulative sum of $N_t$).
2.​ Algebraic Transformation: The equation is rearranged into a quadratic function of $A_t$:​
$$N_t = a + b A_t + c A_t^2$$
3.​ Linear Regression: Run an Ordinary Least Squares (OLS) regression:​
$$N_t \sim A_t + A_t^2$$
○​ The coefficients $a$, $b$, and $c$ are estimated.
4.​ Parameter Conversion: The estimates $\hat{a}$, $\hat{b}$, and $\hat{c}$ are solved
using a system of three equations to find $\hat{M}$, $\hat{P}$, and $\hat{Q}$.

Non-Linear Least Squares (NLS) to Find $P$ and $Q$ (When $M$ is Known)
●​ If the market size $M$ is estimated externally, a standard linear regression cannot be
used directly.
●​ The goal is to find $P$ and $Q$ that minimize the Sum of Squared Errors (SSE):​
$$SSE = \sum (N_t - \hat{N}_t)^2$$
●​ This requires specialized optimization routines (e.g., the optim function in R) to search
for the $P$ and $Q$ values that yield the smallest error.

🤝 Rogers' ACCORD Framework (Qualitative Diffusion Factors)

-​ The ACCORD framework, developed by Everett Rogers, is a qualitative model that
explains why some products or innovations are adopted faster or slower than others,
focusing on characteristics of the product itself.
-​ This framework identifies product-specific characteristics that determine its rate of
adoption:
-​ Bass’s model doesnt have a lot to say for why adoption happens the way it does

A - Relative Advantage

●​ Definition: The degree to which a new product is perceived as being better than the
idea, product, or process it supersedes.
●​ Adoption Effect: Better products are adopted much quicker.
●​ Examples & Caveats:
○​ AI Competition: A superior AI model (e.g., Claude) is expected to be adopted
faster than competitors.
○​ First-Mover Advantage (Caveat): An older product (e.g., ChatGPT) may have
an initial adoption edge due to being first or being forcibly integrated into existing
 ecosystems (e.g., Microsoft Office), giving it an unfair diffusion advantage
regardless of current superiority.

C - Compatibility

●​ Definition: The degree to which an innovation is consistent with the existing values, past
experiences, and current workflow of potential adopters.
●​ Adoption Effect: A technology compatible with existing things is adopted faster.
●​ Examples:
○​ ChatGPT/Search: The landing page is intentionally minimal, looking similar to
the Google search bar, making it compatible with the user's existing online
search behavior.
○​ Google Gemini Integration: Google integrates its AI response directly into the
standard search results page, avoiding making users change their workflow or
go to a separate site.
○​ Robotics Design: Designing a home robot to look like a human makes it more
compatible with the social expectation of having a helper.

C - Perceived Complexity

●​ Definition: The degree to which an innovation is perceived as being difficult to

understand and use.
●​ Adoption Effect: Products that are less complex are adopted faster.
●​ Examples:
○​ Original iPod: Featured a single scroll wheel and minimal buttons, creating the
simplest possible user interface compared to previous multi-button Walkman
players.

O - Observability

●​ Definition: The degree to which the results or use of an innovation are visible to
others.
●​ Adoption Effect: Products that are highly observable are more likely to be imitated and
adopted.
●​ Examples:
○​ High Observability (Fast Adoption): Products that stand out aesthetically
when used in public, like the Tesla Cybertruck.
○​ Low Observability (Slow Adoption): Products used privately, like a new brand
of shampoo, where consumers do not typically discuss or see its use among
friends.

R - Risk
●​ Definition: The perceived chance of loss or negative consequences from adopting the
innovation. Risk is not limited to financial loss.
●​ Adoption Effect: Products perceived as risky are less likely to be adopted.
●​ Types of Risk:
○​ Financial Risk: Losing money if the product is ineffective (e.g., an in-home robot
fails to perform).
○​ Social/Perceptual Risk: Fear of being judged by others for adopting a new trend
(e.g., being one of the first to wear baggy jeans as they come back into fashion).

D - Divisibility (Trialability)

●​ Definition: The degree to which an innovation can be tested or tried on a limited basis
before a full financial commitment is made.
●​ Adoption Effect: Trialability increases the likelihood of adoption by lowering the
commitment hurdle.
●​ Examples:
○​ High-Cost Goods: Test driving a car.
○​ Food/Retail: Offering free samples (e.g., bagels) to allow customers to
experience the product.

🧐 Limitations of ACCORD and the Idea of "The

Chasm"
Incompleteness of the ACCORD Framework

●​ The framework is criticized by economists and marketers for focusing primarily on the
product side and neglecting the consumer side.
●​ Customer Heterogeneity: Adoption timing is often due to differences between people
(heterogeneity), not just product features.
○​ Example: A commuter living far from campus may be an early adopter of an
Electric Vehicle (EV) to save on fuel costs, while a student living close to
campus may have no compelling economic need to adopt early.
●​ Market Evolution: The framework also misses the reality that markets evolve over time:
○​ Price: Prices fall as production becomes more efficient.
○​ Features: Products gain new features via updates (e.g., an AI robot receiving a
software update to enhance functionality).
●​ The ACCORD framework, while useful, is seen as incomplete by many economists and
marketers because it focuses too heavily on the product's attributes and neglects
crucial factors related to the customer and the evolving market.

Customer-Side Limitations (Heterogeneity)

The primary reason people adopt products at different times is because people are different
(customer heterogeneity).

●​ Necessity/Intensity of Need: Individuals have different levels of need or intensity for a

product, driving varied adoption timing.
○​ Example (Electric Vehicle Adoption): A person who commutes a long way
from campus will be an early adopter of an EV to realize significant fuel cost
savings. A person living close to campus, for whom fuel cost is not a major
issue, will be a late adopter. This difference is due to the customers' needs and
circumstances, not just the product's features.
●​ Modeling Implication: This necessity of treating customers differently (understanding
heterogeneity) is why complex demand models (like the Multinomial Logit model in
Week 6) and customer segmentation (Week 3) are essential.

🔄 Market Evolution Limitations

Markets are dynamic and constantly evolve, which affects adoption rates beyond the initial
product design:

●​ Price: As production becomes more efficient, prices for the product typically fall,
opening the door to new market segments.
●​ Features: Products and technologies are continually updated and improved, often
through software.
○​ Example (AI Robot): A customer may buy a "barely functioning" AI robot initially,
knowing that software updates and new features will eventually turn it into a
high-performing tool.

🌉 The Discontinuity of "Crossing the Chasm"

The traditional Bass Diffusion Model assumes a single, smooth S-curve in adoption, moving
seamlessly from low initial adoption to rapid mainstream adoption.

●​ Empirical Finding: Based on real-world observation (popularized in the book Crossing

the Chasm), there appears to be a discontinuity or hurdle between the early adopters
and the mainstream majority.
●​ Adopter Differences: This discontinuity is attributed to fundamental differences in
customer psychology:
○​ Early Adopters: Often not price-sensitive and interested in technology for its
own sake (testing new things).
○​ Mainstream Adopters: Require the product to be proven, reliable, and
integrated (e.g., they need to see others using it successfully).
 ●​ The Chasm: The gap between these two groups is the "chasm"—a point where many
products fail because they cannot transition from attracting technology enthusiasts to
attracting the pragmatic mainstream market. The book is focused on the strategic steps
required to overcome this hurdle and make the product go mainstream
●​

Crossing the Chasm

●​ The BAS model assumes a smooth S-curve from early to mainstream adoption.
●​ The concept of Crossing the Chasm suggests there is a discontinuity or hurdle
between the early adopters and the mainstream majority.
○​ Early Adopters: Often non-price-sensitive, interested in new technology for its
own sake.
○​ Discontinuity: The market must overcome a barrier to achieve mainstream
adoption. The book's focus was on the strategies needed to successfully
navigate this transition.
●​

💲 Customer Lifetime Value (CLTV)

Customer Lifetime Value (CLTV or CLV) is a metric that assesses the total worth of a customer
to the company over the entire period of their relationship.

📝 Definition and Origin

●​ Definition: CLTV is the Net Present Value (NPV) of all future profits expected from a
customer for the duration they remain a paying customer.
●​ Origin: The concept gained popularity with mail-order catalog retailers (e.g., Sears
and Roebuck) in the 1980s.
○​ Costly Acquisition: Printing and mailing large, high-quality catalogs was
expensive, costing several dollars (between $1 and $4 per catalog), resulting in
an upfront loss on every person contacted.
○​ Lifetime Profit: The initial loss was justified by the expectation that the company
would recoup the cost and earn substantial profit over the customer's entire
lifetime through repeated purchases.

CLTV Example: Homelessness Initiatives

An unusual but illustrative example of applying a long-term value perspective:

 ●​ Upfront Cost: Programs in places like Utah provide housing units (costing around
$8,000 per year per person) to people experiencing chronic homelessness. This is
a large upfront expense.
●​ Costly Alternative: The cost of managing homelessness without housing is
extremely high, covering expenses like:
○​ Shelters
○​ Emergency room visits
○​ Ambulance services
○​ Police involvement
○​ Total annual cost estimated at around $20,000 per year per person.
●​ Long-Term Recoupment: By incurring the high upfront cost of housing, the state
recoups the investment quickly by avoiding the much larger, ongoing costs of
crisis management over the individual's time spent housed. This demonstrates
how a short-term loss can lead to a long-term gain (or cost avoidance).
●​ n formal math, CLTV is typically represented by a formula, but in practice, a spreadsheet
is the easiest and most flexible method.
●​ The Compact Formula
●​ CLTV is the sum of the discounted net profit from each period t=1 to T:

●​
●​ Revenue: Revenue generated from the customer in period t
●​ Costs: Costs incurred for the customer in period t (e.g., cost of goods, service).
●​ R: discount rate
●​ T: the time horizon - lifetime
●​ CAC: Customer Acquisition Cost (often an upfront cost, subtracted once)

Recommended Method (Spreadsheet)

The recommended approach is to use a spreadsheet to explicitly track cash flows over time:

1.​ Set up columns for Time Periods (T).

2.​ Input Revenue and Costs for each period.
3.​ Calculate Net Profit ($\text{Revenue} - \text{Cost}$) for each period.
4.​ Discount the Net Profit back to the present value (NPV) using the discount rate $r$
(concept borrowed from Intro to Finance).
5.​ Sum all the discounted profits to get the total CLTV.
●​ The "3X Rule" for Acquisition

When deciding whether to acquire a customer, the financial threshold is an NPV > 0. However,
managers often use a more conservative rule of thumb:

●​ The Rule: Total expected future profits (CLTV) should be 3 times the Customer
Acquisition Cost (CAC).
 ○​ CLTV >= 3 * CAC
○​ Rationale: The 3X rule acts as a risk buffer. Since CLTV relies on
approximations for things like future revenue, churn rates, and the discount rate,
this gap ensures that the acquisition remains profitable even if many of the
underlying assumptions are wrong.
●​ Using CLTV to "Fire" Customers

CLTV is primarily used to identify profitable customers worth acquiring, but it can also be used to
identify unprofitable customers who cost the company money.

●​ Sports Betting Example (DraftKings/FanDuel):

○​ Goal: The betting site wants to acquire consumers who are expected to lose
money gambling over time, making a profit for the site.
○​ Unprofitable Customers: The chart shows that some consumers (on the left
side) are successful and are winning a ton of money from the gambling site.
○​ Action: The gambling site would like to get rid of these successful, unprofitable
customers.
●​ Bank of America Example (Credit Cards):
○​ Initial Loss: The bank gave out free checking accounts to acquire customers,
knowing they incurred an upfront cost (teller salary, software). The hope was to
profit from the interest earned on customer deposits over time.
○​ Unprofitable Expansion: The bank expanded this strategy into credit cards,
granting them too liberally. When customers ran up high debts (e.g., $1,500), the
bank feared they would never pay them back.
○​ Firing Strategy: Bank of America decided to pay customers to close their
high-risk credit card accounts.
■​ Example: Paying the customer $300 to close a card with a $1,500 debt.
The bank accepts the certain loss of $300 to avoid the greater fear of the
customer running up another $1,500 in debt and never paying it back.

​ QUIZZES

Quiz 1:

-​ True or False: A basic data object in R is the [Link] and the Tidyverse set of
packages are designed to work with an enhanced version of the [Link]
called a "tibble": True
-​ Suppose you read in the smartphone customer dataset csv file like so:
-​ cust_dat <-
read_csv("path/to/smartphone_customer_data.csv")
-​ True: calling nrow(cust_dat) or running cust_dat |> summarize(n())
returns the exact same values
-​ geom_point(): Ralph has data for two thousand new social media accounts. For
each account, the data indicate the number of number of posts during the first
 365 days the account was active, and the number of acquired followers during
the same time period. Ralph wants to create a scatterplot of these data to
visualize the relationship between posts and followers. Which ggplot geometry is
most appropriate?
-​ A ggplot template might look something like:
-​ ggplot(data = <DATA>) +
-​ <GEOM_FUNCTION>(mapping = aes(<MAPPINGS>))
-​ where "aes" is shorthand for "aesthetics."In the terminology of ggplot, what is an
aesthetic: The visual properties that can be perceived on the graphic.
Quiz 2:

-​ Which of the following can you use to add a linear regression trend line in a
scatterplot with ggplot2: geom_smooth(method = "lm")
-​ Which of the following options correctly describes what this code returns: A
barplot of the number of customers for each brand, split into separate panels by
gender
-​ ggplot(cust_dat) +​
facet_grid(rows = vars(gender)) +​
geom_bar(aes(x = brand), fill = "dodgerblue4")​
labs(x = "Brand", y = "Customer Count",​
title = "Customer Count",​
subtitle = "by Gender and Brand") +​
theme_bw()
-​ facet_grid(rows = vars(gender)) split into separate panels by gender
-​ True: If you facet on a continuous variable, each unique value of the
continuous variable is treated as a category, so you get one facet per unique
value
-​ True: David is a product manager at a company that rents out high-end
camping gear. The company wants to grow its customer base by targeting
specific outdoor enthusiast segments. David created a segmentation based
on psychological factors such as adventure-seeking levels and
environmental consciousness from a survey of potential customers. He feels
confident about the segmentation because each segment is large and
distinct. David shared his segmentation with his colleague, Sarah.
 -​ Sarah raises concerns that the segments aren't actionable because they can't be
easily identified based on straightforward customer characteristics, like location
or income, which makes it difficult to tailor marketing efforts effectively. True or
False: Sarah's concerns are valid.
-​ Jenna is a marketing analyst at a subscription-based online learning platform.
She created customer segments based on consumer's behaviors, such as how
frequently they engage with courses, the types of subjects they study, and their
subscription plan. Jenna showed her segmentation to her friend Alex.
-​ Alex pointed out that two segments have similar demographic characteristics:
both are mostly made up of young adults between the ages of 25 and 35. He
argues that these segments aren't truly different from each other because their
demographics overlap. Are Alex's concerns valid?
-​ No, Alex’s concerns are not valid because the segmentation is based on
customer behavior.

Quiz 3:

-​ n the R4MRA reading on Principal Component Analysis (PCA), section 8.2.3

involves using PCA for brand ratings.
-​ For the case when ratings are not aggregated (initial case), what is the
minimum number of principal components needed to express at least
84% (or 0.84) of the variance in the data: 6
-​ True: True or False: You should scale the variables to have means equal to zero
and standard deviations of one before running PCA (similar to the scaling we did
before running K-means).
Quiz 4:

-​ Before performing Principal Component Analysis (PCA), why is it important to

convert all variables to the same units (for example, by standardizing them)?
-​ Variables with larger numerical scales would dominate the principal
components
-​ Because PCA is based on variance, variables measured on larger
scales contribute disproportionately more to the total variance.
Standardizing (converting to the same units, e.g., z-scores) ensures
all variables contribute equally to the analysis.
-​ Since all of the variables are of this dataset are in the same units, there is no
need to standardize them before running PCA.
 -​ Run PCA on the 11 score variables. Does the result suggest that there is
only 1 important dimension (ie, "intelligence") or that there may be several
important dimensions (eg, quantitative aptitude, reading/writing, etc.)?
-​ 1 important dimension
-​ Now let's review the multinomial logit model from the Train reading
-​ A researcher gathers data on 342 college students. The first few
students are: Abesh, Britt, Chris, Doug, Ella, Freya, Greg, etc.
-​ For each student, the researcher observes which of the 25 academic
subjects offered by the university the student chooses to major in.
Majors are listed alphabetically:
1.​ Accounting
2.​ Business
3.​ Computer Science
4.​ Dance
5.​ Economics
6.​ Fine Arts
7.​ ... etc.
-​ Also, the researcher collects 9 measures (i.e. variables) of each major that
are believed to correlate with students' choice of major. For example, X1
is the level of mathematical rigor for the major, X2 is the expected salary
for the first year after graduation, etc,
-​ According to the notation in Train's book, what symbols would you use to
represent the probability that Doug chooses to study economics?
-​ Pr (n = 4, j = 5)
-​ What capital English letter represents the total number of alternatives from which
each student can choose: J

Quiz 5:

-​ Fit a multinomial logit model that includes product-specific intercepts and price
as explanatory variables on mdat2 (we called this out4 in the week 5 class
script). After estimating the model, you find that the predicted and actual
product-level market shares are identical. Which feature of this MNL model
specification explains this result?
-​ The use of product-specific intercepts (phone_id).
-​ The Angler Case
-​ Train Chapter 6 presents a case study of angler's preferences for fishing sites.
The empirical work for the case study utilizes a "mixed" logit model.
 -​ Table 6.1 in Train presents the means and standard deviations of the
individual-level parameters (ie, the "random" coefficients).
-​ The image below corresponds to the Campground coefficients, with the mean
marked by a dashed vertical line, zero marked by a solid vertical line, and the
area to the right of zero shaded blue.

-​
-​ What is the mean of the above distribution? (Round to 3 decimal places): 0.116
-​ What is the standard deviation of the above distribution? (Round to 3 decimal
places): 1.655
-​ In the distribution in above, the horizontal axis is for different values of…:
Angler-specific campground preferences

Quiz 6:

-​ In the basic Multinomial Logit (MNL) demand model, why do all customers facing the
same set of options have identical predicted choice probabilities Pnj=Pj:
-​ Because the model assumes identical preference parameters (no
subscript on α and β)
-​ Why can’t we simply add customer attributes (e.g., income, total minutes) directly
to the MNL model to account for heterogeneity?
 -​ They are constant across alternatives and thus cancel out in the choice
probability formula
-​ What is the main purpose of performing k-fold cross-validation?
-​ To assess how well the model’s predictions generalize to unseen
customers
-​ If you use 10-fold cross validation, like we did in Class 6, on models out9 and
out10, you will find that the MSPE (sometimes called just the MSE) is higher for
out10 than out9. What does this mean?
-​ out9 is the better model because its MSPE is lower
-​ Which of the following patterns best indicates that a model is overfitting?
-​ In-sample prediction error decreases, but cross-validated prediction error
increases
-​ Given that out10 has slightly better in-sample fit (higher hit rates and
log-likelihood ratio) but a higher cross-validated MSPE than out9, which
statement best describes their performance?
-​ out9 achieves the best balance between in-sample fit and out-of-sample
predictive accuracy, while out10 adds unnecessary complexity that slightly
harms generalization.
-​

-​ Scenario:
-​ Alex is a pricing strategist at EconoGoods, a company that sells household
essentials. The CMO at EconoGoods is considering revising the pricing strategy
for its flagship product, UltraClean Dish Soap. Alex provides the CMO with
historical monthly sales data and corresponding average prices:

Monthly Sales (Units Average Price per Ounce

Sold) ($)

393 0.853
351 0.921

345 0.946

306 1.002

340 0.949

348 0.943

337 0.978

395 0.884

363 0.912

343 0.971

-​ The CMO assumes that EconoGoods faces a constant price elasticity demand
curve and that the cost to manufacture each ounce of UltraClean Dish Soap is
40 cents.
 -​ What is the price elasticity of demand for UltraClean Dish Soap?​
(This should be a negative number, as a percentage increase in price
should be associated with a percentage decrease in sales.)
-​ Hint: You can calculate the elasticity by regressing the natural
log of quantity on the natural log of price and taking the slope
coefficient estimate. You can do this in Excel, Google Sheet, R, or
wherever you're most comfortable running regressions.​
(Round to 2 decimals)
-​ 1.27
-​ True: True or False: If the CMO assumes a constant elasticity demand curve,
this implies that the percentage change in quantity demanded divided by the
percentage change in price is always equal to the price elasticity of demand, at
any point on the demand curve.

Quiz 7:

-​ What is the economic value to the consumer (ie, their max willingness to pay) to
purchase an otherwise identical Ford truck compared to a Nissan truck, based on
the following conjoint results? In other words, what is the maximal Ford price
premium relative to Nissan, based on the conjoint?
-​ Round to the nearest dollar if necessary, include a negative sign if necessary,
and omit commas and the dollar sign. For example,if the value is negative
$12,345.67 you would enter "-12345.67".
-​

Attribute Level
Part-Worth​
(estimated
coefficient)

Brand Chevrolet -20

Ford 40

Nissan -20

Vehicle Truck 80
Type
 SUV 20

Cross-over -100

Price $50,000 50

$60,000 20

$70,000 -10

$80,000 -40

-​ 20,000

According to the week 7 class R script where we...

●​ Looked at the Samsung phones last year

●​ Noted that the S1 phone was price at $799
●​ Assumed a marginal cost for the S1 phone of $470
●​ Assumed a market size (for the college-aged population) of 10 million
consumers

We began by optimizing the price for the S1 phone in order to optimize the profit of the
S1 phone only. We found that the optimal price for the S1 phone was...?
-​ $709

Then, in the week 7 class R script we

●​ Noted that the price of the S2 phone was $899

●​ Assumed the marginal cost for the S2 phone was $490

We again tested multiple prices for the S1 phone (again holding constant the price of
the S2 phone), but this time we analyzed the total profit to Samsung (from both the S1
and S2 phones). The total-profit maximizing price of the S1 phone was found to be...?
-​ $749
-​ The S1 price that optimized the S1 profit is different from the S1 price that
optimized total Samsung profit. This is because changing the S1 price causes
the estimated probability of buying the S1 and S2 phones to change.
-​ Specifically, compared to the observed prices, a consumer facing a lower S1
price will have an increased probability of buying the S1 phone and a decreased
probability of buying the S2 phone. The estimated probabilities translate into
 estimated market shares, revenues, and profits. So changing the price of any
one phone is likely to affect the profit from all other phones.
-​ Yes
-​ Which of the following quotes are in the Bronnenberg, Dube, Syverson article
from the week 8 reading? Select any or all that apply.
-​
-​ Quote 1:
-​ Brands would likely exist even in the absence of systematic advertising or other
corporate investments in brand-building.
-​
-​ Quote 2:
-​ The total brand value represented by these 100 most valuable brands is
estimated to be $4.14 trillion in 2021 (more than the entire tangible capital stock
of Belgium) and has been growing at an average annual growth rate of 8.1
percent.
-​
-​ Quote 3:
-​ Branding can form associations in the consumer’s memory that assist with recall
and consideration of the branded product.
-​
-​ Quote 4
-​ The economics literature in recent decades has largely overlooked the role of
branding and marketing human capital for our understanding of markets and their
organization as well as firm productivity and macroeconomic growth... Given the
large economic magnitude of intangible brand capital and its recent growth, these
issues seem likely to be of first-order importance.
-​ All of these quotes

Quiz 8:

-​ A dollar today is worth more than a dollar tomorrow, since I can invest that dollar
today and make it worth more tomorrow.
-​ To discount future dollars back into today's dollars (also called the "present
value"), you divide those future dollars by (1+r)^t where r is your discount rate
and t is the number of time periods you need to discount back. You might
familiar with this idea from "discounted cash flows" often discussed in a finance
class.
-​ For example, if you have an 8% annual discount rate and you expect to receive
$100 in 2 years, you compute 100 / (1+0.08)^2 = 85.73.
-​ Often, you have a stream of future payments, say you expect $100 in each of the
next 3 years, then you compute (100 / 1.08^1) + (100 / 1.08^2) + (100 / 1.08^3).
This is often written with summation notation as:
-​ While these equations are compact and look nice in a textbook, the most
common way to do present value calculations is with a spreadsheet:
-​

Tim Profit in Discount Profit in present value (future

e future factor = value times discount
P dollar factor)
e s 1 / (1 + r) ^ t
ri
o
d

0 0 1 0

1 100 1 / 1.08^1 92.59

2 100 1 / 1.08^2 85.73

3 100 1 / 1.08^3 79.38

NPV 257.71

(ie, sum or
"net"
present
value)

-​307.23
-​ Now, you might incur an upfront cost and face attrition.
-​ Suppose you operate a gym. You could pay $1,000 for an advertisement that
reached 100 people. You expect 4% of those people to "convert" into customers
by signing up for the gym. The gym membership is $100 per year, but it costs
you on average $10 per customer per year to service the equipment. In addition,
on average 20% of your customers cancel their membership every year and you
 have a 7% discount rate. What is the NPV over the next 3 years for the
advertisement?

Time Profit in future Discount Profit in

Peri dollars factor today's
od dollars

0 -$1,000 + 4*(100-10) 1 -640

= -640

1 4*(100-10)*(1-0.20)^1 1 / (1.07)^1 = 269

= 288 0.93

2 4*(100-10)*(1-0.20)^2 1 / (1.07)^2 = 201

= 230 0.87

3 4*(100-10)*(1-0.20)^3 1 / (1.07)^3 = 150

= 184 0.82

NPV = -19.14

-​
-​ This negative net present value indicates that the cost exceeds the benefits, and
so it would be a bad advertising investment.
-​ 238,083.83