INTRODUCTION

Forecasting is a common statistical task in business, where it helps inform decisions about scheduling of production,
transportation and personnel, and provides a guide to long-term strategic planning. However, business forecasting is
often done poorly and is frequently confused with planning and goals. They are three different things. Forecasting is
about predicting the future as accurately as possible, given all the information available including historical data and
knowledge of any future events that might impact the forecasts.

Business forecasting has always been one component of running an enterprise. However, forecasting traditionally
was based less on concrete and comprehensive data than on face-to-face meetings and common sense. In recent
years, business forecasting has developed into a much more scientific endeavor, with a host of theories, methods,
and techniques designed for forecasting certain types of data. The development of information technologies and the
Internet propelled this development into overdrive, as companies not only adopted such technologies into their
business practices, but into forecasting schemes as well. In the 2000s, projecting the optimal levels of goods to buy
or products to produce involved sophisticated software and electronic networks that incorporate mounds of data and
advanced mathematical algorithms tailored to a company's particular market conditions and line of business. Besides
the technical or methodological issues that need to be resolved, there are also the human and managerial issues
related to forecasting that need to be taken care of.

Typically there are many methods that can be used to forecast. In our studies we use 5 methods and choose which
one of it can perform the best forecast. The models that we use are naïve with trend model, decomposition method,
simple exponential smoothing model, Holt’s method and Holt-Winters method.

OBJECTIVES OF THE STUDY

The objectives of this study are as follows:

1. To analyze the data into two parts, which is the fitting and the hold-out part.
2. To evaluate the forecasting performance of the models fitted using RMSE and MAPE
3. To identify which of the models perform the best forecast.
4. To generate multi-step-ahead forecast using the best model that already be obtained. 
 Time Series Plot of TOTAL NEWSPAPER SALES (RM)
DATA INFORMATION 3750000

The data consists of newspaper sales from the month

TOTAL NEWSPAPER SALES (RM)

3500000

January 1984 until December 1998, which is t = 180. By

3250000
looking at the pattern, it is a non-stationary series,
because there is an increase in the trend and from the 3000000

firm’s perspective, they would gladly have this (upward)

2750000
kind of series rather than its stationary counterpart. This
is because an increase in sales will add to the profits that 2500000
the firm can acquire. As we know, sales revenue (R) is a
function of quantity of items sold (Q) times price per unit Month Jan Jan Jan Jan Jan Jan Jan Jan
(P). Year 1984 1986 1988 1990 1992 1994 1996 1998

R = P*Q Graph 1
We have divided the data into two parts, which is the fitting part, t = 1-135 and the hold-out parts
t = 136-180. Graph 1 here illustrates the series into a graphical from. In this project, we will be using five types of
models, analyze the hold-out parts and evaluate which model gives the best forecast.

ANALYSIS OF UNIVARIATE MODELLING (FITTING PART)

TOTAL 1) Naïve with trend model

time, t NEWSPAPER ŷ error error squared
SALES (RM), y
1 2426141 2426141.00 0.00 0.00 Naïve model says that what happens today will happen
2 2404993 2384029.34 20963.66 439474987.51 again tomorrow or any other time in the future. This model
3 2446327 2488371.40 -42044.40 1767731310.14
4 2438448 2430594.38 7853.62 61679405.78 is modified to take the trend component into account since
5 2451749 2465122.55 -13373.55 178851918.56
6 2448229 2444714.05 3514.95 12354847.50
some time series data do exhibit the trend pattern, thus
7 2437041 2425904.13 11136.87 124029933.49 becomes the naïve with trend model. It implies that all
8 2439074 2441108.70 -2034.70 4139987.59
9 2472501 2506386.11 -33885.11 1148200683.57
future forecasts can be set to equal the actual observed
10 2505432 2538801.60 -33369.60 1113530523.28 value in the most recent period plus the growth rate, which
11 2524175 2543058.22 -18883.22 356575822.32
12 2498713 2473507.84 25205.16 635300004.01 is the trend value as measured by
13 2533959 2569702.17 -35743.17 1277574069.30
14 2519949 2506016.46 13932.54 194115674.86
Ft+1 = yt*(yt /yt-1).
15 2557022 2594640.41 -37618.41 1415144829.00
16 2555549 2554076.85 1472.15 2167229.93
17 2570534 2585606.87 -15072.87 227191340.84
18 2565276 2560028.76 5247.24 27533578.17
19 2589014 2612971.66 -23957.66 573969549.59
20 2588544 2588074.09 469.91 220819.80
- - - - -
- - - - -
- - - - -
- - - - -
- - - - -
131 3267867 3194243.9 73623.08553 5420358723
132 3404066 3545941.5 -141875.5369 20128667970
133 3341650 3280378.4 61271.55772 3754203785
134 3360543 3379542.8 -18999.81713 360993050.8
135 3281582 3204476.3 77105.69283 5945287867

SSE 699,735,374,630.30
MSE 5,183,224,997.26
RMSE 71,994.62 Table 1: Calculation of Naïve with trend model
MAPE (% ) 1.83
2)
Decomposition method
Decomposition or classical decomposition of time series is a method that describes the behavior of the time series
that can be characterized into four main component types. These are the trend component, the cyclical component,
the seasonal component and the irregular component. Among it, only the trend component and the seasonal
component can be calculated.

Time Series Decomposition Plot for TOTAL NEWSPAPER SALES (RM)

Multiplicative Model
3600000 Variable
Actual MSE = 22,456,200,000
Fits
3400000
TOTAL NEWSPAPER SALES (RM)

Trend

Accuracy Measures RMSE = 149,853.93

3200000
MAPE 4.98068E+00
MAD 1.33187E+05
3000000 MSD 2.24562E+10 MAPE = 4.98

2800000

2600000

2400000

2200000
Month Jan Jan Jan Jan Jan Jan Jan Jan Jan Jan Jan Jan
Year 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995

Graph 2: Trend plus seasonal analysis for decomposition method

As can be seen in Graph 2, from January 1984 newspaper sales start to rise upward until in mid-year 1986. Then, it
starts to move downwards till mid-year 1988. After that, sales of newspapers start showing an upward trend.

The estimated trend would be: Yt = 2,266,847 + 6551*t

3) Single exponential smoothing method

Single exponential smoothing method is the Smoothing Plot for TOTAL NEWSPAPER SALES (RM)
Single Exponential Method
simplest form of model within the family of the
Variable
exponential smoothing techniques. The model 3500000 A ctual
Fits
only requires one parameter that is the smoothing
TOTAL NEWSPAPER SALES (RM)

Smoothing C onstant
constant, α, to generate the fitted values and also 3250000 A lpha 0.488859

the forecasts. This is probably the most popular A ccuracy Measures

MAPE 2
model among the univariate time series modeling 3000000 MAD 45588
technique. The advantage of this technique is that MSD 3953906764

it takes into account the most recent forecast. The

2750000
forecast for the next and all subsequent periods
are determined by adjusting the current period of
2500000
forecast by a portion of difference between the
current forecast and the current actual SSE 533,777,465,673.82
value, which brings to a minimal level MSE 3,953,907,153.14
Month Jan Jan Jan Jan Jan Jan Jan Jan Jan Jan Jan Jan
RMSE 62,880.10
Year 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995
of errors. MAPE (% ) 1.65
α 0.49

Graph 3: Fitted Single Exponential

Smoothing
 Table 2: Results for single exponential smoothing using Solver method.

In determining the initial value, we assume that the actual value is the same as the estimated value, y t = ŷ1. Results in
Table 2 are calculated in Microsoft Excel using the Solver method. There is some difference between the MAPE
obtained in the Minitab software and the Solver method.

4) Holt’s method

The Holt method for autocorrecting an automatic correction method in which the least possible number of data items
is changed and the Fellegi-Holt model is used to determine acceptable sets of values or ranges for the items that are
imputed (OECD website). This technique smoothes the trend and the slope directly by using different smoothing
constants but also provide more flexibility in selecting the rates at which the trend and slopes are tracked. In
determining the initial value, we assume that the actual value is the same as the estimated value, y t = ŷ1.

TOTAL
time, t NEWSPAPER S
SALES (RM), y
Table 3: Holt’s exponential smoothing method using the Solver method
5) Holt-Winters method

Holt-Winters method is a more fined version than the Holt’s method. When seasonality exists, a more suitable
model is needed. Holt-Winter’s trend and seasonality is one such technique that takes into account the trend and
seasonality factors. It consists of three parameters that is the level component, trend component and the seasonality
component. For this project, we assume the series to have an multiplicative effect. In determining the initial value,
we assume that the actual value is the same as the estimated value, y t = ŷ1. The value of the three smoothing
constants is determined by the Solver method.

Winters' Method Plot for TOTAL NEWSPAPER SALES (RM)

Multiplicative Method
Variable
3500000 Actual
MSE = 3,631,746,904
TOTAL NEWSPAPER SALES (RM)

Fits

Smoothing C onstants
3250000 RMSE = 60,263.98
A lpha (lev el) 0.33
Gamma (trend) 0.05
Delta (seasonal) 0.21
3000000 MAPE = 2
Accuracy Measures
MAPE 2
2750000 MAD 45773
MSD 3631746904

2500000

Month n n an a n an an a n a n a n an a n a n
Ja Ja J J J J J J J J J J
Year 4 85 86 87 88 89 90 91 92 93 94 95
8
19 19 19 19 19 19 19 19 19 19 19 19

Graph 4: Fitted Holt-Winters smoothing method

 ANALYSIS OF UNIVARIATE MODELLING (HOLD-OUT PART)

1) Naïve with trend model

TOTAL
time, t NEWSPAPER ŷ error error squared forecast error forecast error squared
SALES (RM), y
136 3445546 3617702.45 -172156.45 29637842945.06
137 3515930 3587751.77 -71821.77 5158366845.36
138 3387084 3262959.74 124124.26 15406833022.17
139 3454797 3523863.69 -69066.69 4770207203.82
140 3384635 3315897.89 68737.11 4724790269.40
141 3411234 3438042.03 -26808.03 718670734.60
142 3395213 3379267.24 15945.76 254267157.32
143 3334954 3275764.49 59189.51 3503398068.60
144 3344788 3354651.00 -9863.00 97278732.86
145 3380577 3416748.94 -36171.94 1308409230.48
146 3240281 3105807.37 134473.63 18083157363.59
147 3397814 3563005.79 -165191.79 27288328948.49
148 3318306 3240658.47 77647.53 6029139324.09
149 3379976 3442792.12 -62816.12 3945865387.21
150 3336736 3294049.17 42686.83 1822165549.51
151 3374515 3412721.74 -38206.74 1459754921.53
152 3336447 3298808.45 37638.55 1416660732.59
153 3354161 3371969.05 -17808.05 317126569.61
154 3388215 3422614.74 -34399.74 1183342261.64
155 3335605 3283811.89 51793.11 2682525841.48
156 3405226 3476300.13 -71074.13 5051532634.86
- - - - - - -
- - - - - - -
- - - - - - -
- - - - - - -
176 3476700 3416321.1 60378.88436 3645609677
177 3642467 3816137.7 -173670.6725 30161502497
178 3509648 3381672.1 127975.8849 16377827128
179 3376224 3247872.3 128351.7068 16474160632
180 3289624 3205245.3 84378.71373 7119767332

SSE 439,939,389,365.47
MSE 9,776,430,874.79
RMSE 98,875.84
MAPE (% ) 2.17

2) Decomposition method

Time Series Decomposition Plot for TOTAL NEWSPAPER SALES (RM) - HO

Multiplicative Model
3700000 Variable
TOTAL NEWSPAPER SALES (RM) - HO

A ctual
Fits
MSE = 6,708,821,061
3600000 Trend

A ccuracy Measures RMSE = 81,907.39

MAPE 2
3500000 MAD 65962
MSD 6708821061 MAPE = 2
3400000

3300000

3200000

Month Apr Oct Apr Oct Apr Oct Apr Oct

Year 1995 1996 1997 1998
3) Single exponential smoothing method

Smoothing Plot for TOTAL NEWSPAPER SALES (RM) - HO

Single Exponential Method
3700000 Variable
TOTAL NEWSPAPER SALES (RM) - HO

A ctual
Fits
MSE = 7,227,696,248
3600000 Smoothing C onstant
A lpha 0.492367 RMSE = 85,015.86
A ccuracy Measures
3500000 MAPE 2 MAPE = 2
MAD 66358
MSD 7227696248
3400000

3300000

3200000

Month Apr Oct Apr Oct Apr Oct Apr Oct

Year 1995 1996 1997 1998

4) Holt’s method

TOTAL
time, t NEWSPAPER S
SALES (RM), y
 5) Holt-Winters method

Winters' Method Plot for TOTAL NEWSPAPER SALES (RM) - HO

Multiplicative Method
3700000 Variable
TOTAL NEWSPAPER SALES (RM) - HO

Actual
Fits
3600000 Smoothing Constants MSE = 4,828,604,070
Alpha (lev el) 0.33
Gamma (trend) 0.05
3500000 Delta (seasonal) 0.21 RMSE = 69,488.16
Accuracy Measures
MA PE 2
3400000 MA D 56571
MAPE = 2
MSD 4828604070

3300000

3200000

Month Apr Oct Apr Oct Apr Oct Apr Oct

Year 1995 1996 1997 1998

SUMMARY OF RESULTS

Model Types

Naïve with Decomposition Single Holt’s Holt-Winters

RMSE Trend Exponential Method method
α = 0.49 α = 0.15 α = 0.33
β=1 β = 0.05
γ = 0.21
Fitted Period :
Jan 1984 – Mar 1985 71,994.62 149,853.93 62,880.10 71,280.13 60,263.98

Evaluation Period :
Apr 1985 – Dec 1998 98,875.84 81,907.39 85,015.86 112,631.31 69,488.16

From the result above, the best model among the five would be the Holt-Winters model, because it has the lowest
RMSE calculated over the evaluation period. The second best model would be the decomposition method, followed
by single exponential smoothing method, naïve with trend model and finally the Holt’s method. In conclusion, the
best model for this series would not be the same if we use some other data series. We can also further study the same
data series by using other models, such as the Box-Jenkins Methodology for example.
MULTI-STEP AHEAD FORECASTS ON THE BEST MODEL (HOLT-WINTERS MODEL)

Winters' Method Plot for TOTAL NEWSPAPER SALES (RM)

Multiplicative Method
4500000 Variable
Actual MSE = 4,853,833,531
Fits
TOTAL NEWSPAPER SALES (RM)

Forecasts
4000000 95.0% PI RMSE = 69,669.46
Smoothing Constants
Alpha (lev el) 0.33
Gamma (trend) 0.05 MAPE = 2
3500000
Delta (seasonal) 0.21

Accuracy Measures
MA PE 2
3000000 MA D 52806
MSD 4853833531

2500000

1 18 36 54 72 90 108 126 144 162 180

Index
REFERENCES

Hyndman, R. J., (2009), Business forecasting methods

Hyndman, R. J., Koehler, A. B., (2005), Another look at measures of forecast accuracy

Lazim, M. A. (2007), Introductory Business Forecasting: A Practical Approach. Second Editon. UPENA

Organization for Economic Co-operation and Development (OECD) website: www.oecd.org