Journal
of Forecasting
5 (1989) 111-116
111
Arnold
L. SWEET
West Lafayette, IN 47907, USA
Purdue University,
Abstract: Additive seasonal models and multiplicative seasonal models can be forecast using general exponential smoothing and Winters methods. The two forecasting methods were compared using 47 of the 1001 time series which were used in the M-competition. Values of the optimal smoothing constants found when fitting the models are shown, Although Winters models always resulted in a better squared error fit, these models gave a better forecast in only 55% of the series. Keywords: Winters. Comparative methods exponential smoothing, Exponential smoothing higher order,
1. Introduction Two methods available for forecasting seasonal time series are general exponential smoothing (GES) using additive and multiplicative seasonal models, and Winters (1960) methods applied to additive and multiplicative seasonal models. These methods appear to be relatively simple to implement, and Winters models have been in use for many years for demand forecasting where distribution inventory management systems are in operation (Gardner (1985)). When examining models of the same type (e.g., additive GES versus additive Winters), both methods will track the same function of time, but study of the variance of the forecast error at different values of the forecast horizon shows that there can be considerable differences in the values that the forecast error can take (see Sweet (1981, 1985), McKenzie (1984a, b)). These differences are of importance in applications, as the variance of the forecast error is used to compute the confidence intervals of forecasts, in various schemes which serve as a warning signal to change the forecast model, and in the
0169-2070/89/$3.50 0 1989, Elsevier Science Publishers
computation of safety stock levels in some inventory control systems. With the above in mind, it was decided to compare the forecasts made by the various models by using some of the 1001 time series which were used in the Makridakis et al. (1982) forecasting competition. It was hoped that one of the methods might prove to be the more accurate most of the time, and so provide some guidance for future applications.
2. Description of the Study The particular series used in this study and their sample sizes are identified in tables 1 through 4 by the numerals used on the data tape, which is available from one of the authors (Carbone (1982)) of the forecasting competition. Only series with quarterly seasonality which had a length of at least 12 times the period were selected in order to minimize the effects of the initialization procedure. There were 23 such quarterly series used.
B.V. (North-Holland)
112
S.M. Bartolo~~i, AL
Table 1 Comparison of the quarterly additive seasonal models. Series no. 191 Sample size 52 Method Constant a GES w 212 217 64 92 GES W GES W 237 286 297 60 60 60 GES W GES W GES W 301 324 56 56 GES W GES W GES W GES W 0.95 0.95 0.30 0.05 0.95 0.50 0.95 0.95 0.95 0.90 0.95 0.80 0.95 0.95 0.95 0.95 0.20 0.90 0.80 0.60 B 0.05 _ 0.05 0.95 0.05 0.05 _ 0.05 0.05 _ 0.05 0.05 0.05 0.05 0.05 0.95 0.95 0.05 0.50 0.95 Y _ 0.20 0.05 Forecasts MSE 3.2E5 lSE6 7416 6378 1.6E5 5.4E5 1.8F.4 1.6E4 465 784 36 238 4.9 4.0 11 3 120 7 9 12 MAPE 5.2 9.3 25.2 20.2 2.6 4.0 13.9 12.8 3.9 5.7 3.5 8.8 1.7 1.5 2.8 1.6 8.4 2.0 2.8 3.2 * * 1 * * * * Best forecast *
326 374
56 56
0.05
* *
The monthly series used had a length of at least 8 times the period, and 24 such series were used. If the length of the data set for a series is N, and the period is p, only the first N - 2p observations were used to fit the models. This procedure is in agreement with that used in the M-competition when p equals 4, and was also used in this study when p was equal to 12 for reasons of consistency. Using a decision process described in the appendix, a choice as to whether to use the ASM or the MSM and values for the smoothing constants were made for both the GES and Winters methods for each series. In order to compare accuracy of the forecasting methods, forecasts were computed at time N - 2p for every time within the horizon of length 2p using both the GES and Winters methods. The mean square error (MSE) and the mean absolute percentage error (MAPE) were computed for the interval of length 2p for every one of the 47 series, which thus allows for a
3. Discussion of Results Of the twenty-tree quarterly series, ten were classified as ASM and thirteen as MSM, while for the twenty-four monthly series, fourteen were classified as ASM and ten as MSM. For all 47 series, both the GES and Winters methods were in agreement on whether the model should be ASM or MSM. In all 47 series, the Winters models provided a better sum of squared errors when fit over the data set than the GES models. Tables 1 through 4 show the optimal values for the constants which were used with each forecast, and the MSE and MAPE for both the GES and Winters methods. Overall, it was found that the Winters models had less forecast error in 26 of the 47 series. It may be of interest to comment on the values
113
multiplicative Method
seasonal Constant a
models. Forecasts P _ 0.05 _ 0.05 _ 0.05 0.05 0.95 0.05 0.05 0.05 0.05 0.05 0.05 0.05 0.05 0.40 _ 0.30 _ 0.10 _ 0.95 _ 0.95 _ 0.95 _ 0.95 0.30 0.40 Y MSE 1192 1494 5.8E6 4.9E6 1432 1215 1321 1121 3537 12497 44 70 97 169 124 7 41 35 38 31 15 36 28 22 6136 3210 MAPE 10.9 12.1 4.8 3.9 19.5 18.0 19.5 18.0 2.5 4.5 4.8 5.7 9.1 11.3 10.3 1.7 3.5 2.9 4.2 3.9 3.0 4.8 2.91 2.86 13.4 9.3 * * Best forecast *
GES W
0.05 0.50 0.10 0.80 0.05 0.30 0.05 0.30 0.05 0.05 0.05 0.40 0.05 0.50 0.05 0.80 0.05 0.70 0.05 0.80 0.05 0.60 0.30 0.90 0.20 0.50
185
100
0.70
189
48
0.60
192
48
0.60
* *
224
56
225
56
GES W
228
56
GES W
238
56
267
60
322
56
* *
336
56
337
56
GES W GES W
375
48
0.95
of the smoothing constants which were found to minimize the sum of squared errors, since such values do not appear to be widely published in the forecasting literature. Examination of the values used for the smoothing constants (shown in tables 1 through 4) showed that when using GES and an ASM the best value of (Ywas at least 0.80 in 13 of the 24 series, whereas when using GES and the MSM, the best value of (Ywas equal to 0.05 in 19 of the 23 series, and was never greater than 0.30. Examination of the results for the ASM when using Winters methods showed that cx equaled 0.95 for five series and 0.05 for five series, whereas j3 equaled 0.05 for 21 series and y equaled 0.95 for three series and 0.05 for 18 series. Similarly, when using Winters methods for the MSM, p equaled
0.05 for 21 series, whereas CY never took values of 0.05 or 0.95. In summary, it appears that the smoothing constants often take values near the ends of the unit interval when using least squares estimates for the models coefficients in the initiation of the forecasting computations.
4. Conclusion Although the Winters models always fit the data better than the GES models, they gave better forecasts than the GES models in only 55% of the series. Gardner (1985) reported on some studies made by comparing the use of GES using trigonometric terms with Winters methods. In one study
S.M. Barioiomei,
Comparison
additive Method
seasonal models. Constant cx Forecasts Best MAPE 6.2 6.8 16.0 23.5 10.7 6.0 6.1 4.8 2.8 5.2 8.0 17.6 7.3 4.2 20.5 14.5 9.6 21.1 9.2 8.9 9.1 19.5 12.2 11.6 5.1 1.1 12.4 29.7 * * * forecast *
Series
no. 389
GES W
0.10 0.05 0.60 0.40 0.80 0.27 0.90 0.05 0.95 0.95 0.60 0.05 0.40 0.38 0.70 0.64 0.50 0.40 0.20 0.17 0.60 0.20 0.10 0.22 0.95 0.40 0.90 0.10
P _
0.95 _ 0.05 _ 0.05 _ 0.05 _ 0.05 _ 0.05 _ 0.05 _ 0.05 _ 0.05 _ 0.05 _ 0.05 _ 0.05 _ 0.20
MSE 14.7 12.7 13.0 28.9 105E5 34E5 21.2E5 11.5E5 399 1151 4.1E5 15.IE5 187E5 9lE5 99185 937E5 382 1402 2lEl 25El 44E7 57El 20284 198E4 19E5 4585 3385 124E5
_
0.05 _
0.05 _ 0.05 _ 0.05 _ 0.80 _ 0.05 _ 0.05 _ 0.05 _ 0.05 _ 0.05 _ 0.05 _ 0.05 _ 0.05 _
391
96
GES W
426
96
GES W
428
96
GES W
* *
447
96
GES W
453
120
GES W
455
120
GES W
458
120
GES W
* *
459
120
GES W
466
120
GES W
* *
415
132
GES W
416
132
GES W
* *
504
108
GES W
520
120
GES W
0.05
0.05
(Reid (1975)), using one-period-ahead forecasting, Winters was better in 72% of the series. There seems to have been no study performed which used the form of the GES models used in this study. Visual examination of the behavior of the series over the last 2p observations in order to observe changes in previous patterns as a possible explanation of the discrepancy yielded no obvious reason for it. Based on the evidence presented above, if the variance of the forecast errors are needed in the desired application, it appears that the GES models would be the preferred choice. Alternately, better accuracy might be obtained
trend
models
Appendix Let x,, t = 1, 2,. . . , N - 2p be the time series observations to be used, and Z*_, be the forecast of x, made at time t-l. Al. Choosing the Best GES Model
For forecasts generated by the GES method, for a fixed value of the smoothing constant (Y,the
S.M. Bartolomei, A.L. Sweet / Exponential Table 4 Comparison of the monthly multiplicative seasonal models. Series no. 387 395 Sample size 120 132 Method Constants QI GES W GES w 427 96 GES W GES W GES W GES W 463 465 461 495 120 120 120 108 GES W GES W GES W GES W 0.05 0.40 0.05 0.40 0.10 0.24 0.05 0.42 0.05 0.53 0.05 0.30 0.05 0.57 0.05 0.64 0.05 0.48 0.05 0.49 P _ 0.05 _ 0.05 0.05 _ 0.05 _ 0.05 _ 0.05 _ 0.05 _ 0.05 _ 0.05 _ 0.25
115
Forecasts Y _ 0.30 _ 0.40 _ 0.41 _ 0.59 _ 0.45 _ 0.50 _ 0.95 _ 0.95 _ 0.86 _ 0.65 MSEX~O-~ 0.16 0.18 9.6 10.3 174 146 2702 779 0.80 0.59 39 47 469 33 68 18 3.1 1.2 691100 23430 MAPE 12.3 13.3 19.2 20.5 11.1 10.4 39.8 17.7 8.1 6.6 15.5 16.4 38.8 8.2 33.7 14.2 23.7 12.8 55.5 9.2
Best forecast * *
434
96
448 449
120 120
* *
* * * *
sum of squared
N-2~
as
dure were obtained by estimating the q and c,s of (2) or the s, and c,s of (3) by minimizing
N--2~
SSE(a)
c t=l
(1)
A search was performed over the interval [0.05, 0.951 for a value of LY which minimized the SSE( CI). The series was then classified as being an ASM or a MSM by choosing the model with the smallest value of min, SSE(a). The models considered when using GES were the ASM, whose deterministic components are given by Y, = 4t + c, > (2)
t=1
c-(xt -Yd2.
The resulting normal equations are linear, and explicit solutions for these coefficients are available in Bartolomei (1985). These values are then used as the initial values in the GES forecasting procedure and also provide a value for J& in (1). A2. Choosing the Best Winters Model Let C-I, j? and y be the smoothing constants which appear, respectively, in the equations for the level, slope and seasonal factors, and let
where q is the slope and c,, t = 1, 2,. . . , p are periodic seasonal coefficients, and the MSM, whose deterministic components are given by Y, = s, + tc, 3 (3)
S=(a,
P,
Y> =
where s, and c, are periodic seasonal coefficients. There are no constraints on the s, and c,s. The initial values used to start the forecasting proce-
c t=l
(x,--,+,)~.
(5)
A search was performed on the cube with sides [0.05, 0.951 for a set of values which minimized the
116
SM.
SSE((U, /3, y). The series was then classified as being an ASM or a MSM by choosing the model with the smallest value of min,,P,u SSE (a, y). j3, When using the Winters method, the models considered were the ASM whose deterministic components are given by Y, = 41 + q2t + b,, (6)
References
Bartolomei, Sonia M., 1985, A comparison of exponential smoothing methods for forecasting seasonal series, Master of Science Thesis (School of Industrial Engineering, Purdue University, West Lafayette, Indiana). Carbone, R., 1982, Computer tape containing the 1001 original time series (McGill University, Faculty of Management, Montreal, Canada). Gardner, E.S., 1985, Exponential smoothing: the state of the art, Journal of Forecasting, 4, l-28. Gardner, ES., and McKenzie, E., 1985, Forecasting trends in time series, Management Science, 31, 1237-1246. Makridakis, S., Anderson, A., Carbone, R., Fildes, R., Hibon, M., Lewandowski, R., Newton, J., Parzen, E., and Winkler, R., 1982, The accuracy of extrapolation (time series) methods; results of a forecasting competition, Journal of Forecasting, 1, 111-153. McKenzie, Ed., 1984a, General exponential smoothing and the equivalent ARMA process, Journal of Forecasting, 3, 333-344. McKenzie, Ed., 1986, Error analysis for Winters additive seasonal forecasting system, International Journal of forecasting, 2, 373-382. Reid, D.J., 1975, A review of short-term projection techniques, in: Gordon, H.A., ed., Practical Aspects of Forecasting (Operational Research Society, London). Sweet, Arnold L., 1981, Adaptive smoothing for forecasting seasonal series, AZZE Transactions, 13, 243-248. Sweet, Arnold L., 1985, Computing the variance of the forecasting error for the Holt-Winters seasonal models, Journal of Forecasting, 4, 235-243. Winters, P.R., 1960, Forecasting sales by exponentially weighted moving averages, Managemeni Science, 6, 324-342.
where the b,s are periodic coefficients constrained so that the sum of the b,s over one period is zero, and the MSM (7) where the sum of the b,s over one period is constrained to be equal to p. When forecasting, renormalization took place after every observation for both of the models. The initial values used to start the forecasting procedure for the ASM was obtained by estimating ql, q2 and the b,s of (6) by minimizing (4). As before, explicit solutions for these coefficients are available in Bartolomei (1985). The model given by (7) is nonlinear in the coefficients ql, q2, and b, and so rather than use a search procedure to estimate these coefficients, it was noted that the model given by (3) is equivalent to the model given by (7) if the ratio of s, to c, were a constant for all t, say s, = kc,, t=1,2 )...)
p. (8)
Thus, the s, and c,s were estimated as discussed above when initializing the GES method and the MSM, and then a mean value of k, denoted by k, is computed from
j=l
where * denotes estimated values. Then ad hoc estimates for use in initializing (7) follow from
P 4^1 'P-l
Biography: Sonia M. BARTOLOMEI holds a B.S. in Industrial Engineering from the University of Puerto Rico (Mayagtiez) and an M.S. in Industrial Engineering from Purdue University. She is currently as assistant professor in the Department of Industrial Engineering Technology at the Regional College of the University of Puerto Rico at Mayagtiez.
c $3
j=l
(10)
and (11) ,...,
Arnold L. SWEET has been Professor of Industrial Engineering at Purdue University since 1980. His interests are in the fields of stochastic processes, time series forecasting, statistical quality control, and applications of probability theory to problems of engineering interest. He is a member of IIE, ASQC, TIMS, and the International Institute of Forecasters.
42 = (w?,,
6, = 4,ij, Further
j=l,2
P.
(12)
(1985).