EXPONENTIAL SMOOTHING In ongoing forecasting systems, forecasts of time series ate made cach period for sueveing periods, Hence the forecasting equation and the estimates ofthe time series ‘parameters need to be updated the end ofeach period to account forthe most resent ‘Observation. This updating must be done to take ito account posible changes in the ‘parameters that may be oceurrng over time. In addition, such changes may imply that ‘unequal weights should be apie to the time srs abscrvtions when the estimates of the parameters are updated. ‘Exponential smoothing is a forecasting method that weights the observed time series values nequally More cet observations are weighted more heavily than more remote ‘observations The unequal weighting is accomplished by using ove or mere smovthing ‘constants, which determine how much weights piven to each observation. Exponential “smoothing has ben found o be most fective when the parameters describing the time series may be changing sowly ove tie. Itis important o note that exponential smocthing ‘methods are not based on any formal statistical model or satsica theory. Rate, these {echniques ae intuitive methods that produce adequate forecasts in some applications. ‘Since these methods have been developed in a piecemeal fashion, no formal statistical rmethodelogy exists for bulding ap exponcntal smoothing mode. In fact, some prac: Uiioners strongly object to using the term model” in the context of exponential smoothing a9380 CuarreR s exronenriat swooTHING We bagin this chapter by explaining simple exponential smoothing in Section 81. Tig ‘tho sssunes tha the time is as 0 rend wile the evel ofthe esas map change slowly overtime. Section 82 discuss adap contol procedure Soc employ what is called racking signal, which canbe sed te tel Us whe &forecating system isnot predicting accurately In such a cae the adaptive conte procedure hs sto take compensatory action to improve the forecsts In Section 8 we present desde exponential smoothing, which is an exponential smoothing method for handing 4 tg seris that displays a slowly changing linear trend. Two approaches are coves: eee parameter double exponential smoothing, which employs tingle smoothing constan\, ted ‘Hole Winters" wo-paranetr double exponential smooth, which employs two sma, ing constants. Section 84 discusses Winters method, which isan expomental approach to forecasting: seasonal data, Both malkipicatie Winters" metho (wht appropriate for inreasing seasonal variation) and additive Wiaters’ method (ebich appropriate for constant seosomal variation) are presented. Section 8S adapts the pr ously discussed methods 1 handling exponential trends and damped trends. Section $6 ‘explain several diferent ways to compute prediction intervals when wsing exponeat smoothing and gives an example of calculating empirical prediction interval Section ‘87 we discuss some of he advantages and disadvantages of wing exponential smoothing methods. We conclude this chapter with optional Section $8, which discusses tre software packages that cau he used o implement exponential smoothing. SIMPLE EXPONENTIAL SMOOTHING Soppose that atime series is appropriately described by the no tend equation neh ‘Wien the parameter remains constant overtime, we have sen that its reasonable o forecast future vals of; hy using regression analysis (ee Example 61), In sucha cae the leas squares point estimate of fy wpe fh hose Sh ‘Wien we compute the point estimate by, we are equally weighting cach of the previously ‘observed time series vale yf, ‘When the value ofthe parameter fi slowly changing over tm, the equal weighting Scheme may uot be appropriate*Tnstead, it may be desirable to weight recent obserme tions more heavily than temote observations. Simple exponential mouthing i afore img method that applies unsqual weights to the time series observations ‘This uneqal weighting ix accomplished by using & smoothing constent that determines how much ‘Sta fama unig oe ne tog be appropri et hi ft byw fy a fa an the ekg thd." A a= snd ee tp xp sea Idi dct i hi Cpt satan aa a etna ete‘ample 81 44) SimMLE exroNeNTIAL swooTHING a eat i attached to each observation. The most resent observation i given the most ight, Oder observations are given successively smaller weights. The procedure allows the forecaster to update the estimate of, 0 that changes inthe value ofthis parameter ‘an he detected and incorporated into the forecasting system, In the following example we illustrate simple exponential smoothing. As tated previc ‘ously, this method should be wed when the model nett describes the time series and when fs may be changing slowly overtime Consider the cod catch data of Example 61, which re given in Table 6.1. The plot of these data (in Figure 62) suggests that the no tend model Kaho he ‘may appropriately desribe the cod cach series. tis alvo possible that the parameter Ay could be slowly changing overtime. ‘We bepin the simple exponential smoothing procedure by calculating an intial est- imate ofthe average lve fy of the series. This estimates denoted (0) and is computed by averaging the fist six time seis values, We obtain E004 a8 boo + 36 “= 6 {Tote tha, since simple exponential smoothing attempts to track changes overtime ia the "average lve! fy by ing newly observed values o update the erates eff me use only six of then ~ 24 time series obvervations to calculate the initial estimate of If we do this hen 8 observation remain 1 tell show fimay be chunaing overtime) xperense dha shown that in general, its reasonable Co calulate inital estimates in exponential smoothing procedure by ting half ofthe historical data. However, team be shown that simple exponential smoothing, wing six observations ie resonable (t would not, a) = 359067 at so diferent from the tue value off thatthe exponential 2 procedure would be adversely alles we ave am estimate 2y(1"— 1) af fy raming tht in tie period T'we obtain 8 new obscrvation js, we can update ag(T ~ 1) 0 a7) which isan estimate made in time period T offi. We compute the ‘pdated estimate by using the smoothing equation og(T) = aye + 8 — sag 1) Here iv smoothing constant between Oand 1 (x willbe discussed in more detail ater} ‘The updating equation say that a(P) the estimate made in time period T of fi. equals a fraction 2 for example, 1) of the newly observed time series observation 9 pos fraction (1 ~ 2) or example, 9 of a(7 — I), the estimate made in tine period T~ 1 of By. The more the average evel ofthe process is changing, the more a newly beret time ‘sie value should influence oor estimate, und thes the large the hooting constant 2 should be set. We will 0a See how to use historial daca To determine an appropriate vale of.2 rane as Couarren s exrontamiac SMOOTHING ‘One: period-ahead forecasting ofthe histories! cod catch time series using singe exponential smoothing with «= 02 a ae ee ‘Acwal Soothes Forsast — coven “Chimma’ —asdutase—Foracant AA Year monn °*%y ay “porod tron’ em =e at) —__tereg_trev___ ere (aio) = 308, ns sa See 38 a : tie Eu xa a Bo te Pu kes 30 oe iy se sae ie a og hme i san = a Sf ie ys oan 309 = ro # Seno 9 3 te x Sow co a ee = ae x tot fe = Sos e = bee a Soe Se, s 2 im ze sae x0 e Be iim Se = ra Be ssi —ee * irae i oe = = 4 rr st = 2 he oe Set Ss 2 of on s a : fe so -ae oe 6 : = ‘We begin with the inital estimate (0) = 38967 and update thi inital estimate y ‘applying the smoothing equation to the 26 observed cod catches. To do thi we wl arbitrarily set 2 equal to 02, and to judge the appropeiatencs ofthis choice of, mew calculate “one periceabead”foreass ofthe historical cod eatchs a8 we catty oat Se smoothing procedure. Since the atl estate of fy is a(0) = 35967, follows hat 38 ‘isthe rounded forecast made tine Ofer y,the vale of he time src in peso 1. Sia ‘we Se from Table 6.1 that y = 342, we have a forcast eeor of 362 ~ 360 ~ 2 Uiag 1 = 362, we cam update 29() td), am estimate made in period 1 ofthe averagele ‘ofthe time series, by using the equation ag(t)= 9, + (= 2)0y0) = 02060) +.98(388.67) = 35972 Since this implies that 340i he rounded forecast made i period 1 for J. and sine see from Table 6 that y ~ 381, we have a forecast error of 360 — 381 = 21. Us| sig 381, we ean update d(1) 10-43} an estimate made in period 2 of fy by winramese fy simpue exponenTAL soorHMe ‘equation (2) ay + (1— adattd = G81) + 98099.72) an Since this implies that 360s the rounded forecast made in period 2 or ys, and since we fee fom Table 6 that y, —317, we have a forecast eror of 317 ~ 360 = ~83. This procedure is continued through the entire 24 periods of historical data. The results are marized in Table 8.1 Using ce results this able, we find that, for 2 =.02 the sum ‘Ot uqunted forecast erors is 27,74, To Sind a “good” valu os, we evaluate the sum of Sauared forecast erors for valics of ranging from.02 30 in inerements of 02 (a most ‘ponent! smoothing applications, the value ofthe smoothing constant use is between ‘and. 30), Noting thatthe results are given in Table 82, we find that = = 2 mininnes ithe sum ofquaed forecast errors This small value of «indicates thatthe average level fol the ime series not changing much overtime. “The sume of squared forcast rrors for different values of 8 Sree Be Bsr it eo a Be Doe gore gots [9 ? 2 ae a we & ‘Bom 3 mate {In genera, simple exponential smoothing i earied out as flows: ‘Simple Exponential Smoothing ——— 1. Suppose thatthe de series yy deseribed by the model neh ‘where the average level fy may be slowly changing aver time, Thea the ‘Mmateop(T) of fy made in time period T's given bythe smoothing ‘equation fT) ayy + (L ~ ada = 1)7 ‘uarren s BxpONENTIAL SMOOTHING ‘where is.a smoothing constant between Oand 1 and ay(T'— 1) the ‘timate of fy made in time period T 2A point forecast made in time pesiod fry. aT +N ayes +0 TAT) + bynes T+ nal aces ea a i pd ah [al T) + Hoy 280 Einar where A(T) =" — = A tae obey n bie T+ eo oT | Neen erat asec Ar+ = “Therefore, a point forecast made i ie period T+ TOF Yeysoeis | PravdT + Dal + 1) ad a 100(1 ~ 2)% prediction interval made in time period T +1 for Dreseis hasacr +0 Faampiesd In Example 8.1 we saw tat 2 = 02 “good” vale of the smoothing constant wie, TRncautg the 24 sere od cates ia Tate 61. Therefore, we will se spt ‘Shonen smoothing with «= 02 forecast future moutly cod catches. From Tt SPE that (4) 3863 the timate ade in month 24 ofthe average ee fy | Titus that the pot forcast made in month 24 of the cod al in month 25 aad tn) oer oture monthly cod ctchis Saan(24) = a(24) = 39618 Moreover, since Sy, —aute= 0 A (+ p= aol "4 Wa ~ 429) 7m | _ p= smay.epn = 3970 4-4 n= 90 2» ~ a0| 1) source bxPoneNTIAL SMOOTHING 14.95% prediction imerval made in month 24 for ys it [386.13 + sjoan.2SN(28)] = [38613 4 1960.25)028:62)] = (28601,42625) ‘Now, assuming that we observe a co catch in January of year 3 of yoy = 364, we ean update u(24 10 (25) by using the equation (25) = ayas + (1 ~ 9028) = 024388) +.98(356.13) = 35669 “This impli thatthe point forecast made ia month 25 ofthe cod catch im month 26 nd ‘fany cer future mostly cod catch is Saye) ~ o25) = 39609 Frthermore we ean update ACM) 1 AS) by wing he equation 24024) + yn4 ~ a4) ag = 208 tl MH (_ 7282) 4/384 ~ 386131 Des — Bs = 2859 It otlows that a 95%, petition interval made in onth 25 fo aes [38669 & sony 2580251] = [9689 + (1.96911. 25128-9)) = (28664,42674) 1m genera, note thatthe smoothing equation ag(T) = ayy + (0 ~ shag T = 0 implies that ag = 1) ayy. +0 ~ adadT = 2) Substitution, therefore, gives vs y(T) = apy + (1 ~aNfay, +1 ~ ada — 2) = ayy tal ays (aa —2) Subottutingrecusively for ag(T ~ 2) g(T ~ 3). dll vand sf, we obtain al T) = apy + al 9)fras tall AP yea tt allay + = a) 00O) ere wee that (7 the etiate mae in ive pri Toth average eve fy ofthe tine seriey, can be expressed in terms of the observations Js Youn, Frm sooinital estimate, (0). The coeents measuring the contributions of the of Ye Pent Yeas ne Ye— tbat 5, a2), a(t — 2, = a)" decreate en ponentaly with age. For this reason we refer to this procedure as simple exponen ‘smoothing ‘ince the coefficients measuring the contributions of yy, ty Yrs» Yet dy creasing exponentially, the most recent observation), makes the largest conicibution ‘he current caimate of f. Older observations make smaller and smaller contbation ‘this estimate, Thus remote observations are “dampened out” of the curent estima. ‘of fea time advances. The ate at which remote observations are dampened out depends ‘ou the smoothing constant a. For values of near I, emote observations are dampevel ‘out quickly. For example fa =.9, we obtain cocciens 9, 08,009, 0009, 0 Fy Values of «near 0, remote observations are dampened out more slowiy (2 i, we ‘ade by simulated forecasting of ‘Simulation suggest that a value ofthe smoothing constant greater than. 3 wold bebe, ‘we should consider using other methods (o be presente i ter chapters to forecast te {ime series. In such a case the mode! woh te tay not he appropiate o the parameter fy may be changing guickly over ime Inte ‘cave the use of simple exponential smoothing is probably not hes. 8.2 ADAPTIVE CONTROL PROCEDURES tis sometimes necessary to change the smoothing constants) being employed inex ‘ncotial smoothing For example, when simple exponential smoothing is being wed Be tae at which fs changing over time migh change. This might suggest that a diet ‘smoothing constant & would produce improved forecasts. Decisions about changing ‘the smoothing constant can be made by employing what i called an adaptive com! procedure ‘We start by discussing a racking signal which belps us to decide when something wrong with 2 forecasting system (lor instance, when we are using an inappropee smoothing constant). Although we know that a forecasting system will never prods peroct forccasts, a tacking signal can be wed to tell us when our forecast errors wt Jaeger than an “accurate” forecasting system might reasonably be expecied vo produ “Tobegin with suppose that we havea history of Tsingle-period-ahea forecast er 6,62) 6,0 ey Here a) denotes the particular valve of «employed to obtain ‘Ingle period ahend forecast errors. We next dfine the following sum (of thes fo vanaf