Analyzing Data:

Indexing Through Data:

- LEFT(Cell, # of figures (inclusive)). Example: LEFT(19970401, 4) = 1997
- RIGHT(Cell, # of figures (inclusive)). Example: RIGHT(1997, 2) = 97
- NUMBERVALUE(text). To convert text to numbers
- CONCATENATE(Cell1, Cell2, Cell3) to combine
- YEAR(Cell), MONTH(Cell), DAY(Cell)

Returns and Dividends:

- Holding period return = R = (P1+D)/P
- No dividend return = P1/P
- (Holding period return – no dividend return)*P = Dividend

Summary Statistics:
- AVERAGE(range of cells)
- Average Daily: (1+AVG)^(255) -1
- MEDIAN(range of cells)
- SKEW(range of cells) – 0 if symmetrical, positive is right, negative if left
- STDEV(range of cells)
- Day STDEV into Year = STDEV*SQRT(255)
- MIN()
- MAX()
- Plotting Histogram
o Calculate Min and Max
o Construct Bins through Min + X(interval is up to you)
o Plot = COUNTIF(range of cells, “>”&BIN). This will give you total counts of numbers greater than
the X tick
o Calculate changes between the plot points, then divide changes by total count to get
probability that it shows up
- If statements
o COUNTIF(columns to count, condition if met)
o If you’re calculate probability like (how often WMT’s return was below -5%), make sure to use
COUNTIF(range, “>-1000”) if there is -99999 invalid data
o Others: SUMIF, AVERAGEIF
- Logical Statements: returns TRUE (1) or FALSE (0)
o = (4<5) = TRUE, (4 = 5) = FALSE
o AND(test1, test2)
o OR(test1, test2)
o IF(criteria to be true, if true, if false)
o Ohers: COUNTIFS, SUMIFS, AVERAGEIFS (multiple conditions)
- Ex. AVERAGEIFS(Range to average, Range1, condition1, range 2, condition 2)
- Correlation:
o Corr = COV[X,Y]/)StdDev[X]*StdDev[Y] where COV[X,Y] = E[X*Y]-E[E]*E[Y]
o -1 < corr < 1
o Excel: CORREL(list1, list2)
- Odds Ratio: Another way of measuring co-movement. Compare unconditional probability to probability
conditional on some fact (When Market was bad for 17 days, WMT was bad for)
o Calculate how many bad days market had vs WMT
 o Conclusion should be: Market having a very bad day increases the odds of WMT having a very
bad day by a factor of 60!
Regressions:
- If we assume the relationship is linear, what is our best guess of Yi conditional on knowing Xi

- Error: everything about the Y that is not explained by the X

- Alpha and beta tell us about the relationship between X and Y. if Beta > 0, then higher X is associated
with higher Y
- Choose alpha and beta to minimize the sum of squared errors
- Regression to the Mean: Things tend to go back to mean over time
- R^2 measures how much of Y is explained by X, aka how strong the relationship is
- Standard error (sB) tells us how widely the true distribution of B(beta) varies
o Approx. 68% of distribution falls within one standard deviation
o Smaller Sb → precise estimate (68% of distribution falls within one Sb)
o T-stat = B/sB (beta/standard error of beta)
- T-State > 2 = <5% chance the true relationship is zero (above zero is relationship)
- R^2 VS T-Stat: R^2 doesn’t take into account the sample of the data, only how well they fit
- Excel:
o Use INTERCEPT and SLOPE
- INTERCEPT(Column of Y’s, Columns of X’s) = to get alpha
- SLOPE(Column of Y’s, Columns of X’s) = to get beta
o Use Excel’s Analysis Toolpak
o Use LINEST and INDEX functions together
- Allows multivariate regression, provides plenty of info
- LINEST(column of dependent variables (Y),
column of independent variables (X), 1,1)
• Third 1’s is for if you want constant
Alpha. 0 if you want to force Alpha = 0
• 4th argument is 0 if you just want to
output the coefficient estimates, and 1
if you want additional info
• 3 and 4 will always be set to 1
• Use INDEX(LINEST, row, column)
o Ex. Standard Error of Beta = INDEX(LINEST, 2,1)
o Ex. Alpha = INDEX(LINEST, 1, 2)
- Example: Estimate CAPM alpha and Beta
o CAPM is E[R] = Rf + B*Rp, so we regress R[asset] – Rf on Rm – Rf
- Slope of this equation will be Beta
- DO NOT DO: Stationary vs Non-Stationary Variables
o Non-stationary: Increases or decreases without bond (GDP, Stock Price, Population)
o Stationary: Varies around mean but tends to stay in the same range (growth rate), or ratios, or
moving average (smoother line, take average over time)
- Autocorrelation and EMH
o Autocorrelation is correlation between variable Xt and its lagged value Xt-1
o It is also equivalent to the slope of regressing Xt on Xt-1: Xt = alpha + B*Xt-1 + Et
o If autocorrelation is different from zero, then Xt+1 is predictable by Xt: the best estimate of
Xt+1 is alpha + B*Xt
 o However, efficient markets (risk adjusted) future returns should not be predictable by publicly
available information
- Multivariate Regression
o Just as easy to have multivariate regression: Y linearly depends on X1 and X2
- Make sure when Linest is used to make the columns adjacent to each other

o
o Make sure you normalize variables. Example, use police spending per capita, instead of just
police spending.
o Only BETA tells you about relationship of something. Alpha doesn’t
- Control Variables
o Reverse Causality: Maybe it’s the Y variable that is driving the X
o Confounding Variable: Maybe there is a third variable Z that is driving both X and Y →
something deeper.
- Example: Police Spending and Crime
o First regression is done with violent crimes and police spending, you get a pretty good
correlation: But do more police spending cause more violent crimes? That could be a reverse
causality
o Second regression is with Justice Expenditure AND Past Violence, and you can see that there is
negative slope with Justice Expenditures, meaning that I actually somewhat reduces violent
crime (however this was insignificant looking at t-stat). This is controlling for regular crime
- Note: important that you use the right control variables
- Survivorship Bias
o Example: Suppose you want to estimate average return on a set of stocks

-
- Looking at this, you would average 100% over time.
- However, you need to account that firms with -100% probably went bankrupt
- Hence when you average across all firms that exist today, you get 150%
o Dealing with it: Follow all the firms through, even if they’ve been knocked off follow them
anyway
o Peso Problem and Equity Premium Problem: Even though historical US equity premium is 6%
with 18% volatility, why isn’t everyone investing in equity?
- This is because expected US equity return going forward is not 6%, it is because US is a
survivor
- Other markets like Russia, Germany and Argentina were much worse
- Peso Problem: infrequent shocks happen that may significantly affect stock prices, but
they are rarely observed in financial data
Valuation:
Methods:
- DCF: PV of firm’s expected FCF

o Gordon Growth Method:

- One may have higher multiple than other because it has higher expected growth g or
lower discount rate r
- Multiple Valuation: Find comps, and compute multiples

- Balance Sheet Review

o Total Assets: Total Liabilities + Equity = Unlevered Assets + PV of Tax Shields
o Total CF = Liability CF + Equity CF = Unlevered CF + Interest Tax Shield CF

o
- Tax Shields
o Depreciation tax shield: D&A is non-cash expense so add it back after taxes because no money
was spent. Any money spent to pay for D&A is in capex
o ITS: Deduct int when computing tax bill, when computer total CF add int back because int isn’t
lost, it is paid to one of the firm’s stakeholders. Don’t include in equity CF. Ignore int when
computing UFCF
- Valuation Approaches
o CFTE: Valuing equity by discounting CF
to Equity by rE. Downside is that you
need to project entire debt schedule to
compute FCFtE
o WACC: Value assets by discounting UFCF
by rWACC. Value equity by subtracting
liabilities from assets. If debt is a held at
constant ratio, WACC is easiest.
o APV: Value unlevered assets by
discounting UFCF by rU.
- Value ITS by discounting ITS CF
(use either rD or rU depending
on how debt is structured).
- Value assets as sum of unlevered
assets and ITS.
 - Value equity by subtracting debt from assets. When debt is constant or schedule is
known in advance, APV is easiest to use. Discount tax shields by rE
- DCF
o Terminal value approach is better than raw GGM, because in TV you forecast CF up to date T
where the firm should be stable.
o The terminal value approach can be used with FCFtE, WACC or APV

o
- CF and Balance Sheet Flow
o NWCt = NWC + change in NWC
o PPEt = PPE + capex – D&A
o If you claim too much depreciation, you’ll need to pay additional taxes
o Capital employed = PPE + NWC
- Estimating Growth Rates:
o For stable firms, you can estimate by averaging over historical growth rate. For
o For growing/declining firms/industries, make an assumption about industries long term growth
rate (in 5 years), and how cash flows will grow until then
o 4 Methods:
- Naïve Method: Use historical average
- Alternative 1: Use analyst forecasts
• Sometimes not reliable because internal biases, or they have worse models
- Alternative 2: Use Return on Capital (ROC) and Re-Investment Rate (RIR)
• Re-investment Rate= (Net Cap Ex + change in NWC)/(EBIT(1-t))
o Net capex is capex – depr. Tells us now many dollars of new capital
created per dollar of EBIT
• Return on Capital = EBIT(1-T)/(BV Debt + BV Equity). How many dollars of EBIT
are created per dollar of capital?
• G = (return on capital)*re-investment rate
o (Capex – D&A + Change in NWC)/ (BV Debt + BV Equity)
• Works well if firm is stable and you have good estimates
- Alternative 3: Relate growth to observables (i.e CapEx and R&D), build statistical model
o Usually you would want to use historical US GDP growth, as you probably won’t grow any faster
than that in the long run
o * When calculating growth, (last year/this year), only use positive numbers for denominator or
else you’ll get negative growth even when its positive (example, -2 to 1)
- Example: Valuing Microsoft
o Calculated FCFtE through the reg equation, but write AND statement saying that it should only
be calculated if not -99999. The year should also be +1 from first year available, because you
can’t compute changes prior
o Calculating Growth Rates:
- Naïve Method:
• Calculated the average EBITDA/EBIT, RE, FCF and FCFE YoY change, but the
numbers are huge because we threw out negative numbers
o EBITDA/EBIT are less likely to be negative, but FCF’s are
o 1 year firm-level earnings growth is too noisy and not useful
o Do 5 year growth instead. Ex. (Average t1, t+5)/(Average T-2, T-5)
 ▪ Once we did this the numbers were much more consistent
o Rule: Use 5 year growth rather than 1 year growth
- Alternative 3: Statistical Model
• Use CapEx to Total Assets and R&D to total assets as predictors of EBITDA
growth
o Start with does capex to total asset in t predict EBITDA growth in t to t+1
▪ Hypothesis: Firms that invest more will grow faster

▪
o Copy over EBITDA for t+1 and t, and align it next to Capex and TA -1.
o Filter to remove missing data
o When using 1 year capex to EBITDA, it was way too noisy
o Next: Use past 3 years (add past 3 years) to predict next 5 years EBITDA
growth (EBITDAt+5/EBITDAt)

o
o Even though t-stat is significant at 5%, r^2 is still small, because total
growth is a combination of aggregate growth and growth relative to
other firms (and this is hard to forecast). Another reason is because
EBTIDA is sometimes really small (outliers) so we should remove.
o However, the results weren’t strong because we predicted EBITDA
t+5/EBITDA, and we threw out all negative EBITDA’s. Instead, we should
use (EBITDAt+1 – EBITDA)/ TA
o
- Estimating Discount Rate R for Equity
o Depends on valuation approach you use: Can use
- Historical equity return
- Asset pricing model (CAPM or Fama-French 3 factor) to estimate equity return
o Method 1: Historical Equity Return:
- Take average (monthly/year/daily), but this isn’t good estimate because they won’t
grow at that rate forever
o Method 2: Use CAPM
- Regress Firm Return - rF/ Market return – rF
- Use INTERCEPT/SLOPE to get alpha and beta
o Method 3: Fama French 3-Factor Model
- Multivariate regression → use LINEST function

-
o *If a firm has leverage then typically
- Estimating Discount Rate R for Debt
o Calculate cost of debt over weight averaged of different issues of debt
- Estimating WACC
o Use market cap and long term debt (comes from BS, book/market usually similar)
o Do WACC calculation
- Getting WACC to rU
o For APV, you need rU
 o
- Putting it all together
o Don’t use GGM blindly unless firm is stable
o Estimate FCF, then calculate TV, and discount it

- Comps Analysis

o
o Numerator is market value, and denominator are measuring fundamentals
o Steps:
- Identify firms in same busines as the firm you want to value
- Calculate multiples for comps and come up with an estimate of the multiples for the
firm you want to value (often take average or median of comps multiples)
• ME/BE  market value/book value
- Multiply the estimate multiple be the BE of the firm you want to value
- Enterprise-level (all shareholders) multiples better than equity multiples
- Pros:
• Incorporate a lot of info about other valuations
• Embodies market consensus about discount rates and growth rates, free-ride on
market’s information
• Provide discipline in valuation process by providing benchmark
- Cons:
• Implicitly assumes all companies are alike in growth rates, cost of capital, and
business composition. Hard to find “true” comparable firms
• Hard time incorporating firm-specific information
• Book values can vary across firms depending on the age of a company
• Accounting differences in calculating the multiples
• If market is wrong, you will also be wrong
- Using Multiples
• Important to keep firms as similar as possible on as many dimensions as possible
• One important difference between firms that may appear similar is leverage
o Hard to compare firms with diff financial lev
o Should never have something related to equity in numerator and to total
firm in denominator
• To account for leverage
o Use EV or asset-based multiples, or use an equity multiple and an
adjustment for leverage derived from M&M (don’t do) or
o Use characteristics regression approach
 •
• Gamma1 = 1/(r-g) and gamma0 = 0

•
• Higher the FCF higher the V, so to find V you do = g0 + g1*FCF + E
- Regressions
• V = B0 + Bat*Assets + Blt*Liabilities
• Benefits of regressions over multiples is that any number of characteristics can
affect a firm’s value
o Multiples only assumes one characteristic (the denominator)
o

Diff in Diff:

Hypothesis: Do tax cuts lead to high growth? Do tax cuts pay for themselves?
- Positive View:
o Donald Trump: Biggest tax cut in history, will lead to real GDP increases of 1%
o Paul Ryan: Tax Bill is pro growth, will help grow
- Negative View:
o Fitch: May lead to short term boost to output, but will not pay for themselves
o Congressional Budget Office: Federal deficit will rise a lot
- Principles:
o Lower taxes: better incentives for people to work, and corporations to produce
o Lower taxes: gov’t has less resources so unable to provide essentials = bad for economy

Method:
- Using a Diff-in-Diff
o First difference: Compute the difference in growth before and after the action
▪ Can compare this to multiple episodes in history (time series difference)
o Second difference: Compare first difference to a second subject without the action
▪ Negative diff-in-diff = action that happened was negative
▪ Positive diff-in-diff = action that happened was positive
Pre – Action Post- Action Diff
 Subject 1 (Action) =avg(time before) =avg(time after) =post – pre
Subject 2 (No Action) =avg(time before) =avg(time after) = post – pre
Diff = no action – action B =no action- action (A) = A- B
o Caveats:
▪ Too simple to be taken seriously
▪ Needs to be repeated on many episodes
▪ Need to find good comparables
▪ Needs to choose correct dates and timing

- Using a Regression
o Line up all the data points in one long list, with your X and Y variables, then run the regression
o This is called a pooled regression because every (t,i) data point is treated like any other,
regardless of what time or state it comes from
o Controls:
▪ Ideally you want to control for various other characteristics (population, density.etc
▪ Essentially, try to keep everything else constant
o Endogeneity:
▪ We usually care about causation not just correlation, and depending on the scenario,
both variables may be caused by a difference force
• Reverse Causality: example: fans sell more when more firefights die
- Other concepts
o Experiments:
▪ Treatment group is one the experiment is applied on, control is not
▪ In economics, experiments are usually impossible
▪ There can be natural experiments that happen (Kansas for example)

Application:
- Kansas:
o Approach 1: For each tax change, compute a diff-in-diff. The computer average diff-in-diff
across all tax increases, and tax cuts
▪ Problem: not all tax cuts are equal in size
o Approach 2: Regression: Regress change in GDP growth around time of tax change on the size
of the tax change
▪ Run 3 regressions, one with tax cuts and one with tax hikes cause it’ll be more accurate
▪ Across all states and time periods, regress the change in GDP or state i in period t on the
change in taxes of state I in period t
▪ Regress change in tax (t, t-1)
▪ Regress change in GDP (3 years ahead – 3 years behind, average(T,T+2 ) – average(T-3,
T-1)
• Used because GDP YoY is too noisy

Predicting Bankruptcy
- Probit Regression:
o Default is yes/no event. WE will assign variable Dt,I a value of 1 is firm defaults in time t, and 0
otherwise
o
o Creating lead and lag variables: We may want to forecast default in the future. We can create
the same row for Dt+1 or Dt+2.etc
▪ Line up data for X variable the same, but for data for Y variable, you can do it T years
ahead