Forecast 4or <1 ())). $('#$'+ 1# ; #' Accurac- Measures MAPE. 1,#'& MA/. (,#*' M!E. 1)&)#(1 Forecast 4or <1 ())). $$,#)+) c# =oo>ing at the error measures3 there is not much di44erence 1et6een the t6o choices o4 smoothing constant# he error measures 4or ? ; #, are slightl- 1etter# he 4orecasts 4or the t6o choices o4 smoothing constant are also not much di44erent# d# he residual autocorrelations 4or ? ; #, are sho6n 1elo6# he residual autocorrelations 4or ? ; #' are similar# here are signi4icant residual autocorrelations at lags 13 , and 5ver- nearl-7 &# A 4orecasting method that -ields no signi4icant residual autocorrelations 6ould 1e desira1le# , 15. A time series plot o4 @uarterl- Revenues and the autocorrelation 4unction sho6 that the data are seasonal 6ith a trend# A4ter some experimentation3 Winters multiplicative smoothing 6ith smoothing constants ? 5level7 ; )#&3 A 5trend7 ; )#1 and B 5seasonal7 ; )#1 is used to 4orecast 4uture Revenues# !ee plot 1elo6# Accurac- Measures MAPE $#& MA/ '"#1 M!E 111,'#, Forecasts <uarter Forecast =o6er :pper +1 (,,,#'$ ((+*#$, ('1$#"( +( 1"&+#"& 1++$#&, (()(#1( +$ (($+#"& 1"'"#($ (*)'#+( +, 1&&+#+, 1**"#,' ((1'#)1 +* (,*'#1& ()'*#+) (&,'#'* +' 1""+#$' 1*,$#1) (,*1#'( *
An examination o4 the autocorrelation coe44icients 4or the residuals 4rom Winters multiplicative smoothing sho6n 1elo6 indicates that none o4 them are signi4icantl- di44erent 4rom Cero#
17. a# he 4our-6ee> moving average seems to represent the data a little 1etter# Dompare the error measures 4or the 4our-6ee> moving average in the 4igure 1elo6 6ith the 4ive-6ee> moving average results in Figure ,-,# ' 1# !imple exponential smoothing 6ith a smoothing constant o4 ? ; #+ does a 1etter Eo1 o4 smoothing the data than a 4our-6ee> moving average as Eudged 1- the uni4orml- smaller error measures sho6n in the plot 1elo6#
19. a# he results o4 Folts smoothing 6ith ? 5level7 ; #" and A 5trend7 ; #1 4or !outh6est Airlines @uarterl- income are sho6n 1elo6# A plot o4 the residual autocorrelation 4unction 4ollo6s# It appears as i4 Folts procedure represents the data 6ell 1ut the residual autocorrelations have signi4icant spi>es at the seasonal + lags o4 , and & suggesting a seasonal component is not captured 1- Folts method# 1# Winters multiplicative smoothing 6ith ? ; A ; B ;#( 6as applied to the @uarterl- income data and the results are sho6n in the plot 1elo6# he 4orecasts 4or the 4our @uarters o4 ())) are. <uarter Forecast ," &&#"') *) 1&,#&11 & *1 1&1#,', *( 11+#"&*
he 4orecasts seem reasona1le 1ut the residual autocorrelation 4unction 1elo6 has a signi4icant spi>e at lag 1# !o although Winters procedure captures the trend and seasonalit-3 there is still some association in consecutive o1servations not accounted 4or 1- Winters method# " CASE 4-1: THE SOLAR ALTERNATIVE COMPANY his case provides the student 6ith an opportunit- to deal 6ith a 4re@uent real 6orld pro1lem. small data sets# A plot o4 the t6o -ears o4 data sho6s 1oth an up6ard trend and seasonal pattern# he 4orecasting model that is selected must do an accurate Eo1 4or at least three months into the 4uture# Averaging methods are not appropriate 4or this data set 1ecause the- do not 6or> 6hen data has a trend3 seasonalit-3 or some other s-stematic pattern# Moving average models tend to smooth out the seasonal pattern o4 the data instead o4 ma>ing use o4 it to 4orecast# A naive model that ta>es into account 1oth the trend and the seasonalit- o4 the data might 6or># !ince the seasonal pattern appears to 1e strong3 a good 4orecast might ta>e the same value it did in the corresponding month one -ear ago or 8tG1
; 8t-11 # Fo6ever3 as it stands3 this 4orecast ignores the trend# Hne approach to estimate trend is to calculate the increase 4rom each month in ())* to the same month in ())'# As an example3 the increase 4rom 9anuar-3 ())* to 9anuar-3 ())' is e@ual to 581$
- 817 ; 51+ - *7 ; 1(# A4ter the increases 4or all 1( months are calculated3 the- can 1e summed and then divided 1- 1(# he 4orecast 4or each month o4 ())+ could then 1e calculated as the value 4or the same month in ())' plus the average increase 4or each o4 the 1( months 4rom ())* to the same month in ())'# Donse@uentl-3 the 4orecast 4or 9anuar-3 ())+ is 8(*
; 1+ G I51+ - *7 G 51, - '7 G 5() - 1)7 G 5($ - 1$7 G 5$) - 1&7 G 5$& - 1*7 G 5,, - ($7 G 5,1 - ('7 G 5$$ - (17 G 5($ - 1*7 G 5(' - 1(7 G 51+ - 1,7JK1( 8(*
; 1+ G 1( 1,& ; 1+ G 1( ; (" he 4orecasts 4or ())+ are. 9an (" Fe1 (' Mar $( Apr $* Ma- ,( 9un *) 9ul *' Aug *$ !ep ,* Hct $* Lov $& /ec (" Winters multiplicative method 6ith smoothing constants ? ; #13 A ; #13 B ; #$ seems to represent the data 4airl- 6ell 5see plot 1elo67 and produces the 4orecasts. Month Forecast 9anK())+ 1"#& Fe1K())+ 1&#) MarK())+ ('#& 1) AprK())+ $(#) Ma-K())+ ,(#, 9unK())+ ,*#& 9ulK())+ *&#, AugK())+ *&#" !epK())+ ,+#' HctK())+ $$#+ LovK())+ $$#* /ecK())+ (&#)
he na0ve 4orecasts are not unreasona1le 1ut the Winters 4orecasts seem to have captured the seasonal pattern a little 1etter3 particularl- 4or the 4irst $ months o4 the -ear# Lotice that i4 the trend and seasonal pattern are strong3 Winters smoothing procedure can 6or> 6ell even 6ith onl- t6o -ears o4 monthl- data# CASE 4-2: MR TUX his case sho6s ho6 several exponential smoothing methods can 1e applied to the Mr# ux data# 9ohn Mos1- tries simple exponential smoothing and exponential smoothing 6ith adEustments 4or trend and seasonal 4actors3 along 6ith a three-month moving average# !tudents can 1egin to see that several 4orecasting methods are t-picall- tried 6hen an important varia1le must 1e 4orecast# !ome method o4 comparing them must 1e used3 such as the three accurac- methods discussed in this case# !tudents should 1e as>ed their opinions o4 9ohnMs progress in his 4orecasting e44orts given these accurac- values# It should 1e apparent to most that the degree o4 accurac- achieved is not su44icient and that 4urther stud- is needed# !tudents should 1e reminded that the- are loo>ing at actual data3 and that the pro1lems 4aced 1- 9ohn Mos1- reall- occurred# 11 1# H4 the methods attempted3 Winters multiplicative smoothing 6as the 1est method 9ohn 4ound# Each 4orecast 6as t-picall- o44 1- a1out (*3&(*# he error in each 4orecast 6as a1out ((N o4 the value o4 the varia1le 1eing 4orecast# $# 9ohn should examine plots o4 the residuals and the residual autocorrelations# I4 Winters procedure is ade@uate3 the residuals should appear to 1e random# In addition3 9ohn can examine the 4orecasts 4or the next 1( months to see i4 the- appear to 1e reasona1le# CASE 4-3: CONSUMER CREDIT COUNSELING 1# !tudents should realiCe immediatel- that simpl- using the 1asic naive approach o4 using last period to predict this period 6ill not allo6 4or 4orecasts 4or the rest o4 1""$# !ince the autocorrelation coe44icients presented in Dase $-$ indicate some seasonalit-3 a naive model using April 1""( to predict April 1""$3 Ma- 1""( to predict Ma- 1""$ and so 4orth might 1e tried# his approach produces the error measures MA/ ; ($#$" M!E ; &'1#$, MAPE ; 1&#"* over the data region3 and are not particularl- attractive given the magnitudes o4 the ne6 client num1ers# $ !ince the data have a seasonal component3 Winters multiplicative smoothing procedure 6ith smoothing constants ? ; A ; B ;#( 6as tried# For these choices. MA/ ; 1"#("3 M!E ; *,*#,1 and MAPE ; 1'#+,# For smoothing constants ? ; #*3 A ; B ; #13 MA/ ; 1'#",3 M!E ; ,*1#(' and MAPE ; 1,#$)#
CASE 4-4: MURPHY BROTHERS FURNITURE 1# Lo ade@uate smoothing model 6as 4oundO A Winters multiplicative model using ? ; #$3 A ; #( and B ; #1 6as deemed the 1est 1ut there 6as still some signi4icant residual autocorrelation# $# Pased on the 4orecasting methods tested3 actual Murph- Prothers sales data should 1e used# A plot o4 the results 4or the 1est Winters procedure 4ollo6s#
1$ An examination o4 the autocorrelation coe44icients 4or the residuals 4rom this Winters model sho6n 1elo6 indicates that none o4 them are signi4icantl- di44erent 4rom Cero#
Fo6ever3 9ulie decided to use the na0ve model 1ecause it 6as ver- simple and she could explain it to her 4ather# CASE 4-5: FIVE-YEAR REVENUE PROJECTION FOR DOWNTOWN RADIOLOGY his case is designed to emphasiCe the use o4 su1Eective pro1a1ilit- estimates in a 4orecasting situation# he methodolog- used to generate revenue 4orecasts is 1oth appropriate and accuratel- emplo-ed# he >e- to ans6ering the @uestion concerning the accurac- o4 the proEections hinges on the accurac- o4 the assumptions made and estimates used# Examination o4 the report indicates that the anal-sts 6ere conservative each time the- made an assumption or computed an estimate# his is pro1a1l- one o4 the maEor reasons 6h- the Pro4essional Mar>eting Associates 5PMA7 4orecast is considera1l- lo6er# !ince 6e do not >no6 ho6 the accountant proEected the num1er o4 procedures3 it is di44icult to determine 6h- his revenue proEections 6ere higher# Fo6ever3 it is reasona1le to assume that his 4orecast o4 the num1er o4 cases 4or each t-pe o4 procedure 6as not nearl- as sophisticated or thorough as PMAs# here4ore3 the recommendation to management should indicate that the PMA 4orecast3 6hile pro1a1l- on the conservative side3 is more li>el- to 1e accurate# /o6nto6n Radiolog- evidentl- agreed 6ith PMAMs 4orecast# he- decided not to purchase a "3&)) series D scanner# he- also decided to purchase a less expensive MRI# Finall-3 the- decided to o1tain outside 4unding and did not resort to an- t-pe o4 pu1lic o44ering# he- 1uilt their ne6 imaging center3 purchased an MRI and have created a ver- success4ul imaging center# CASE 4-: WEB RETAILER 1# he time series plot 4or Hrders sho6s a slight up6ard trend and a seasonal pattern 6ith pea>s in /ecem1er# Pecause o4 the relativel- small data set3 the autocorrelations are onl- computed 4or a limited num1er o4 lags3 ' in this case# Donse@uentl- 6ith monthl- data3 the seasonalit- does not sho6 up in the autocorrelation 4unction# here 1, is signi4icant positive autocorrelation at lag 13 so Hrders in consecutive months are correlated# he time series plot 4or DPH sho6s a do6n6ard trend 1ut a seasonal component is not readil- apparent# here is signi4icant positive autocorrelation at lag 1 and the autocorrelations die out relativel- slo6l-# he DPH series is nonstationar- and o1servations in consecutive time periods are correlated# $# !imple exponential smoothing 6ith ? ; #++ 5the optimal ? in Minita17 represents the the DPH data 6ell 1ut3 li>e an- QaveragingR procedure3 produces 4lat-line 4orecasts# Forecasts o4 DPH 4or the next , months are. Month Forecast =o6er :pper 9ulK())$ )#1),* )#)+&+ )#1$)$ AugK())$ )#1),* )#)+&+ )#1$)$ !epK())$ )#1),* )#)+&+ )#1$)$ HctK())$ )#1),* )#)+&+ )#1$)$ he results 4or simple exponential smoothing are pictured 1elo6# here are no signi4icant residual autocorrelations 5see plot 1elo67#
1*
*# It seems reasona1le to 4orecast Dontacts directl- i4 the data are availa1le# Multipl-ing a 4orecast o4 Hrders 1- a 4orecast o4 DPH to get a 4orecast o4 Dontacts has the potential 4or introducing additional error 5uncertaint-7 into the process# CASE 4-!: SOUTHWEST MEDICAL CENTER 1# Autocorrelation 4unction 4or total visits suggests time series is nonstationar- 5since autocorrelations slo6 to die out7 and seasonal 5relativel- large autocorrelation at lag 1(7# $# I4 another 4orecasting method can ade@uatel- account 4or the autocorrelation in the otal Visits data3 it is li>el- to produce Q1etterR 4orecasts# his issue is explored in su1se@uent cases# CASE 4-": SURTIDO COO#IES 1# 9ame learned that !urtido Doo>ie sales have a strong seasonal pattern 5sales are relativel- high during the last t6o months o4 the -ear3 lo6 during the spring7 6ith ver- little3 i4 an-3 trend 5see Dase $-*7# $# Winters multiplicative smoothing 6ith ? ; A ; B ; #( seems to represent the data 4airl- 6ell and produce reasona1le 4orecasts 5see plot 1elo67# Fo6ever3 there is still some signi4icant residual autocorrelation at lo6 lags# 1' Month Forecast =o6er :pper 9unK())$ '*$(*, "1$*1 1(1*1*+ 9ulK())$ +1(1*" 1,1,*$ 1(&(&'* AugK())$ '**&&" +*$'& 1($',11 !epK())$ 1*$(",' ",1',+ (1(,(,* HctK())$ 1+1)*() 11)+*$$ ($1$*)+ LovK())$ (1$$&&& 1*1&$*, (+,",(1 /ecK())$ 1")$*&" 1(+,+)( (*$(,+'