Scribd Logo
Upload

Welcome to Scribd!

  • Time Series Analysis BA

    Uploaded by

    Ravi Sista
    24 views54 pages

    Copyright

    © © All Rights Reserved

    Available Formats

    PDF or read online from Scribd

    Did you find this document useful?

    Is this content inappropriate?

    Report this Document

    Copyright:

    © All Rights Reserved
    Download as PDF or read online from Scribd
    Flag for inappropriate content
    24 views54 pages

    Time Series Analysis BA

    Uploaded by

    Ravi Sista

    Copyright:

    © All Rights Reserved
    Download as PDF or read online from Scribd
    Flag for inappropriate content
    of 54
    Time Series Analysis and Forecasting co: 18.1 TIME SERIES PATTERNS 184 LINEAR TREND PROM CHON, Horizontal Patern IS$_ SEASONALITY Trend Pattern Seasonality Without Trl sonal Pattern : ‘ svonality with Tend Wap iRciahe na) Moaels Bases an Mots Data Cyclical Pate Selecting a Fo APPENDIX 15.1 FORECASTING WHILENCEL DATA ANALYSIS TOOLS ing Method! Scanned with CamScanner 734 Chepter 15 Time Series Analysis ond Forecasting » The purpose of this chapter is to Provide an introduction to time series analysis and fore. Casting, Suppose we are asked to provide quarterly forecasts of sales for one of our com. Pany's products over the coming ‘one-year period. Production schedules, raw materials ’ Purchasing. inventory policies, and sales quotas will all be affected by the quarterly fore. Casls We provide. Consequently, poor forecasts mav evant inn TT Toe 7 Covts for the company. How should w, 20 about providing the quarterly sales forecasts” Goal judstent intuition. and an awareness of the state of the ecouomy may give us ough idea or “feeling” of wh 's ikely to happen in the future, ht converting that feeling into a nurnber that can be use a next year’s sales forecast is challenging, The Manage cy iiemt Science in Action, Furecasing Energy Needs in the Utility Industry, describes. the A forecasts ier rots rrr le that forecasting plays inthe utility industry, inthe fare. Mt Forevastin thas ca be clastfied as qualitative or quantitative. Qualitative meth must cept ae fhe tev innge vil a ale 5 valve the we of expert judgment to develop forecasts Such methoale ate Shoronmate when historical data on the variable being forecast petfesttrees inplicable Quantitative onecaving metho can be used sf 01) past formation abut the variable hei information cx ble to assume tht past is prologue te the pattern oft the taure) Ne will focus excasivey on quate forecasting methods inthis chapter the historical data rericted o past alues ofthe veutlonste forecast the fore ‘esting procedure is called sime series method and the historical data are referred to as a FORECASTING ENERGY NEEDS IN TI E UTILITY INDUSTRY* Duke Energy iva diverted enery company with The largest observed electric demand for any sare tual gas and clecui businesses given period. such av ee 2 day. a month, srvany_ to 2006, a year, is defined as the peak lod. The forcag wilh Cinergy of Cincinnati, of the annual electric pea Tee a Nn America’s largest decision for comets fa EnetE) Companies with asets ong moe than ad the finance 270 bilo As atest of this mevser the Cin: Obviously. tiny decision that feads ty hav annati Gas & Electric Company: became pat of 8 the wait available mo sooner tha necessary i Duke Enctgy, Talay: Duke Energy services oer crucial 85 null seta ele North Carolina, South C fndiana, and 0» Fores ides the timing hetating units, pact ofthis decision 4s great Cand coors The enry foe anger ina Ohi. Kents, ways she Fen so, Canad forthe generating wn rs nth win es same eb a en) aca The ean et amas estes tate ene chet prt one we ts et om fe peed htm fasted nah em the in sexo: th planing frat change sat rece sl tt ag d dxivion-mahing pe ns” This stuson plies cee ean inpwtne manage mul nla fnecod Ont asin conguy er = man cg andi sa dson maken vst he Te ea ton atl acs ig based on the forecast Lancs requir ety Ploeticl shortages ate not just branineias” or “bla i arriving at these Products M0, untibe the decisions. . tut fm,» great amount ¢ cepsrog an he coven om the fo psc: “Bawd a8 informatie onde by by R sity Eins of Duke Enerzy a all Scanned with CamScanner 15.1 Time Series Pateens time series. The nbjective of ime series analysis is to uncoxer a pattern in the historical data or time series and then extrapwlate the pattern inte the future; the forecast is haved solely on past values ofthe variable andor» past forecast etrons In Section 15.1 we discuss the various hinds of time series that a forecast faced with in practice, These include a ul pattern, a trend, pattem, both a trend and a seasonal pattern, and a eyetical patern. In ender to build a (quantitative forecast measurement of forecast ac ‘curacy. Different measurements of forecast aceurae advantages and disadvantages. are discussed in Section 15.2 In Section 15.3 we consider the simplest ‘which isa horizontal or constant patter, For this pattern, we develop the classical moving average, weighted moving average. and exponential smoothing inodels. Many time series have a trend, and taking this trend into account is important: in Section 15.4 we provide regression models for finding the best model parameters when a linear tend is present Finally, in Section 15.5 we show how to incorporate both a trend and seasonality into a forecasting model might be » seasonal fonstant of horizon noel tis alo necessary to have nd their respect 15.1 TIME SERIES PATTERNS: ‘A time series is a sequence of observations on a variable measured at successive points in time or over successive periods of time. The measurements may be taken every hour, day, week, mont, o year, or at any other regular interval! The patern of the data is an Jmportant factor in understanding how the time series has behaved inthe past. If such be~ havior can be expected to continue in the future, we can use it to guide us in selecting an appropriate forecasting method. ‘To identify the underlying pattern in the data, 2 useful frst step is to construct atime series plot. A time series plot is a graphical presentation of the relationship between time ‘and the time series variable; time i represented on the horizontal axis and values of the time series variable are shown on the vertical axis. Let us first review some of the common types of data patterns tat can be identified when examining a time series plot. Horizontal Pattern ‘Ahorizontal pattern exists when the data uctuate randomly around a constant mean over time, To illustrate a time series with a horizontal patern, consider the 12 weeks of data ‘Table 15.1, These data show the number of gallons of gasoline (in 1000s) sold by @ gasoline distributor in Bennington, Vermont, over the past 12 weeks. The average value ‘or mean for this time series is 19.25 or 19.250 gallons per week. Figure 15.1 shows a time series plot for these data. Note how the data fluctuate around the sample mean of 19,250 gallons. Although random variability is present, we would say that these data follow a horizontal pattern “The term stationary time series? is used to denote a time series whose statistical prop- erties are independent of time. In particular this means that 1. The process generating the data has a constant mean. 2. The variability ofthe time series is constant overtime “Web our discs tine series or wich eves che aris re acoded ot euainierals, Cases in which the chert ore mode ot vequol itera or byod the ope of is es ora feroal dln of sfonoriy, se . Ord ord Rides (2012, icp of Busines Forecasting Moxon, OH Cengage Leoring, p. 155. Scanned with CamScanner 736 CChopter 15 Time Series Analysis ond Forecasting TABLE 15.1 GASOLINE SALES TIME SERIES Week Sales (1000s of gallons) 0 1 19 ’ 2B 16 16 2» 18 2 20 5 2 RRM FIGURE 15.1 GASOLINE SALES TIME SERIES PLOT —_—-. ————uwuwuU wesley l 2 3 4 5 Gasoline 6 7 8 9 0 1 2 8 2 g Eo a A time series plot fora stationary time series will always exhibit «horizontal pats random fluctuations. However, simply observing a horizontal pattem is not sufisien ‘evidence to conclude that the time series is stationary. More advanced texts on forecasting discuss procedures for determining if atime series is stationary and provide methods transforming a time series that is nonstationary into a stationary series. ‘Changes in business conditions often result in a time series with a horizontal tem that shifts toa new level at some point in time. For instance. suppose the g2col distributor signs a contract with the Vermont Sate Police to provide gas police cars located in southern Vermont beginning in Week 13. a Scanned with CamScanner TALE V5.2 GAYOLINE GASB TIME WNIT IE VEIMONT Werk Sales (1H of galls) Week ' 7 a 2 a 3 wo oi f Cnrioatovond 5 % % 6 6 0 7 w % 4 % 1 9 n y x] Mal” Sales (100s of gallons) 3 : : WPT SAS 67 BO IM IDI IS 16 1718 19 25.21 22.23 2 Week jE ae no es et asc the distributor naturally expects to see a substantial increase in weekly sales starting in Week 13. Table 15.2 shows the number of gallons of gasoline sold forthe original time series and the 19 weeks after signing the new contract. Figure 15.2 shows the cor- responding time series plot. Note the increased level ofthe time series beginning in ‘Week 13. This change im the level ofthe time series makes it more dificult to choose an appropriate forecasting method, Selecting a forecasting method that adapts well to changes in the level of atime series is an important consideration in many practical applications. Scanned with CamScanner 738 wal Bice Chopter 15. Time Series Analysis ond Forecasting Trend Pattern Although time series data generally exhibit random fluctuations, a time series may also show gradual shifts or movements to relatively higher or lower values over a longer period of time. Ifa time series plot exhibits this type of behavior, we say that a trend pattern Crist. A trend is usually the result of long-term factors such as population increases or decrease, sifting demographic characteristics of the population, improving technology, and/or changes in consumer preferences, To illustrate a time series with a of bicycle sales for a paticular manufacturer over the past 10 Table 153 and Figure 15.3. Note tat 21,600 bicycles were sold in Year 1, 22.900 linear trend pattern, consider the time series 6 shown in TABLE 15.3 BICYCLE SALES TIME SERIES Year Sales (1000s) 216 2.9 255 219 239 25 315 297 286 314 1 FIGURE 15.3. BICYCLE SALES TIME seers PLOT. (SNOITIIN S) LO7d SAPYAS SNL SANAATY OAUG TONALSTIOHD ¥'S1 FUN 5 2 8 NIATY NVA TOULSTIOHD P'S VL omiseu scanned with Camscanner air sor es 1 apy aaK LE ‘SaIVAS ANIL SAIVS VTTEWANN S'S TeV Euys020104 puo ssfpouy sung ou ¢| s¥do4>) siegun Fs gam scanned with Camscanner souas aun atp ut wasaad ag Sew reap stoagg fe24f2%9 yun [Pap TOU op ax sardey> stp ul “sI2aY9 afa4o-puan se o1 pouta}a4 yf earn “yNsa4 SY "SERIO uysnoy 40j 242s au pea ue uoneyur pides yo spotad Xq panoyjoy uoueyut resapous Jo spouad "ajdwexs ‘sojoKo ssoutsng seaKninus 01 anp st sauas aun & Jo ywauodwod [3 ppuan atp axoge pure mojaq suotessasqo Jo sts seynau Yi JOLAEOG 98 au qusouoDa Aueyy se94 au wey 2xOu 30 SIs ENN BUH] PUA aKP AXOGE PU HOI swutod Jo 2ouanbas Sunewsaye we swoys yoyd sauas own 2p Jt Sstxo waWEd [e>H>%> Y usayeg peat “Aiyeuoseas pue puan 4109 Yi ‘ajqedeo s1 ip poxpout Supseaas0yv asm 0 paou ax s2se9 Yons Ul '9pPS Uo} ‘osqe wsaued [euose2s yetp apnou09 x ‘sy, f pu ¢ sso4senb ul soqsty pu seaK Y>ed JO sayrenb puooas ayy ut sasto] a soyes tein sareaqpul ose gC aan V9AINOH “Wuasaud St pues Surseouout ue S293 “smeak sno} ysed a4) 4940 Jauntgegnuc so|MaqHEN B20) S3peS 195 a1 uauenb woys 9°51 aundty ut rod souas own Suypuodsa.i09aip pur gS SIAEL 1 ejep 24 “2ouR|su! Jo, “UHoHed yeuOseas F PUE Puan E HOG apo}ouSALDS aU wos waayyeg [euoseag pue purl], -quasaid st waned euoseas Couenb e yeyp apnjouod pnom am “sniy,“auunpon sajes samo] 94 BAK OF spuay s2yenb yuunoy ayp pu “sajes 1sausiy ayt sey Jauenb puooas a4 ‘sajes avesapour ‘iy siourenb pany pue wsy 24) St yey, ak yoea UIYHHN ssn990 rey EIep ay) UL UsOME neuorsKs esjeanas o[d souias aun atp ut suontentony yo uonsadsuts950)9 YoHamOH Catt Ta Poway aus, KAMA HAAN EEL OS HET 0 scanned with Camscanner CChopter 15. Time Series Anclysis ond Forecasting TABLE 15.6 QUAR Seeeneess seen Year Quarter Sales (1000s) 1 1 48 2 4 ’ 3 60 4 65 1 58 2 = 3 68 4 14 3 1 6.0 i 3 4 2 3 4 56 15 18 63 59 80 84 _—.] FIGURE 15.6 QUARTERLY TELEVISION SET SALES TIME SERIES PLOT Quarterly Television Set Sales (1000s) (SS Se ee es agate a) 4eSee6 yi 8) 9) (eta 12 19: 1451516 Period eee Selecting a Forecasting Method ‘The underlying patiem in the time series is an important factor in selecting a forecasting method, Thus. atime series plot should be one of the frst analytic tools employed when trying 1 determine which forecasting method to use. If we see a horizontal patter, then we need to select a method appropriate for this type of pattem. Similarly, if we observe a tend Scanned with CamScanner 15.1. Time Sevies Patterns in the da 74 then We ned to use a feasting method that is capable of handling a trend effectively, Inthe next to sections we illustrate methods for assessing forecast accuracy and consider forecast ‘Management Science in Action, Products, describes th for its consumer and ofice products, ruealels that can be used in situations for which the underlying pattem is horizontal; in other words, ao trend oF sease consider methods appropriate when trend and/or season 1 effects are present. We then ity are present in the data. The ‘recasting Demand for a Broad Product Line of Office considerations made by ACCO Brands when forecasting demand FORECASTING DEMAND FOR A BROAD PRODUCT LINE OF OFFICE PRODUCTS* ACCO Brands Corporation is one of the work's largest suppliers of branded oflice and con- sumer products and print finishing solutions. The company’s widely recognized brands in clude AT-A-GLANCE®, Day-Timer*, Five Star’, GBC*, Hilkoy*, Kensington*, Marbig, Mead*, NOBO, Quartet*, Rerel, Swingline’, Tiibrat Wilson Jones*, and many others, ‘Because it produces and markets a wide array ‘of products with a myriad of demand characteris- ties, ACCO Brands relies heavily on sales forecasts in planning its manufacturing, distribution, and marketing activities. By viewing is relationship in terms ofa supply chain, ACCO Brands and is cus- tomers (which are generally retail chains) establish close collaborative relationships and consider each other to be valued partners. AS a result, ACCO Brands’ customers share valuable information and data that serve as inputs into ACO Brands fore- casting process. ner role as a forecasting manager for ACCO Brands, Vanessa Baker appreciates the importance of this additional Information. “We do separate forecasts of demand for each major custome” sid Baker, “and we generally use twenty-four to thity- six months of history to generate monthly forecasts twelve to eighteen months into the future. While trends are important, several of our major product Jines, including school, planning and organizing, and decorative calendars, ae heavily seasonal, and sea- sonal sales make up the bulk of our annual volume.” Daniel Marks, one of several account-level strategic forecast managers for ACCO Brands, adds: 3 The supply chain process includes the total lead tine from identifying opportunites to making or procuring the prodoct to geting the product onthe shelves to align withthe forecasted demand this can tently take several months, so the accuracy of cur forecast is cftcal throughout each step of the supply chain, Adging to this challenge is the mish ‘of obsolescence We sell many dated items, such as planners and calendars, which have a natural, built in obsolescence. In ation, many of cur products ‘ature designs that are fashion-consciows or contain pop culture images, and these products can also, be- come obsolete very quickly as tastes and popular- ity change. An overly optimistic forecast for these ‘rodoets can be very costly; but an overly pessis tic forecast can result in lost sales potential and give ‘ur competitors an opportunity to take market share fromus. Inadktion to looking at trends, seasonal com- ponents, and cyclical pattems, Baker and Marks ust contend with several other factors. Baker notes, “We have to adjust our forecast for upcom- ing promotions by our customers” Marks agrees and adds ‘We also have ogo beyond just forecasting consumer demand; we must consider the retiler’s specific needs in our order forecasts, suchas what typeof display wil be used and how many units of pra uct must ben display to say their presentation requirements. Curent inventory is anathe fatr a customer searing ether too moch o 0 file invertor, that wl et thee future orders, and we need lo reflect that in ou forcast. Wil the praast have a short life because itis ed Toa cultural fad? ‘Whatarethe retailer's marketing and markdown sa tees? Our knowledge ofthe environmen in which ‘our suply chain panes are competing helps us to forest demand more accurately, and that races ‘east and makes our cstomes, a well as ACCO Bau, far more prot *The authors are indebied to Vanessa Baker and Daniel Marks of ACCO Brands for providing input forthis Management Science in Acton Scanned with CamScanner 744 Chapter 15. Time Series Ancyss and Forecasting 15.2 FORECAST ACCURAC is section we begin by developing forecasts for the gasoline time series shown in Tae $1 ving the simplest of all the forecasting methods, an approach that uses the most ‘seen ek’ es volume asthe frees forte met week, For instance, the distributor ~ sold 17,000 gallons of gasoline in Week I this value i use as the forecast for Week 2 Next, we use 21 the aul value of sales in Week 2, ashe forecast for Week 3, and soon, The forecasts obtained forthe historical data using this method are shown in Table 15.7 in the column labeled Forecast. Because of its iplicity, this method is often referred to as naive forecasting method. How accurate are the forecasts obtained using this naive forecasting method? T ‘wer this question we will introduce several measures of forecast accuracy. These measures are used to determine how well a particular forecasting method is able to reproduce the lime series data that are already available. By selecting the method that is most accurate for the data already known, we hope to increase the likelihood that we will obtain more accurate forecasts for futute time periods, ‘The key concept associated with measuring forecast accuracy is forecast error. If we denote ¥, and F, as the actual and forcasted values ofthe time series for period r, respec- tively, the forecasting error for period tis NSS 4 aha (15.1) {hat is, the forecast error for ime period 1s he diference between the actual and the forecasted values for period t TABLE 15.7 COMPUTING FORECASTS AND MEASURES OF FORECAST: ACCURACY USING THE MOST RECENT VALUE AS THE FORECAST FOR THE NEXT PERIOD Absolute Absolute Time Value of Squared Value of Series Forecast Forecast Forecast Percentage Percentage Week Value Forecast Error Error Errors Error Evron T u z 21 7 4 4 16 19.05 19.05 3019 2 2 4-053 1053 2 oe 4 4 16 1739 1739 5 18 3B 5 5 ar 78 6 16 18 2 2 4 12.50 1. 16 4 4 16 20.00 8 8° "2 2 4 1 9 2 18 4 4 16 18.18 Hoe 0 2p "2 2 4 10.00 Noo mM Hs 5 5 233 2 2 45 a 7 49 31.82 Total 3 1 211.69 Scanned with CamScanner 15.2. Forecast Accuracy 745 For instance, because the distributor actualy sold 21,000 gallons of gasoline in Week ‘Zand the forecast using the sales volume in Week 1, was 17,000 gallons, the forecast error in Week 2is Forecast Error in Week 2 The fact that the forecast error is positive indicates that in Week 2 the forecasting method underestimated the actual value of sales. Next we use 21, the actual value of sales in Week 2, asthe forecast for Week 3. Since the actual value of sales in Week 3 is 19, the forecast eror for Week 3 is ¢s = 19 — 21 = ~2. In this case, the negative | forecast error indicates the forecast overestimated the actual value for Week 3. Thus, ! the forecast error may be positive or negative, depending on whether the forecast is 100 low or too high. A complete summary of the forecast errors for this naive forecasting method is shown in Table 15.7 in the column labeled Forecast Error. It is important to note that because we are using a past value of the time series to produce a forecast for period 1, we do not have sufficient data to produce a naive forecast for the first week of this time series. ‘A simple measure of forecast accuracy isthe mean or average of the forecast errors. If wwe have m periods in our time series and kis the numberof periods atthe beginning of the time series for which we cannot produce a nave forecast, the mean forecast error (MEE) is | MFE = St (52) SUM EEE Table 15.7 shows that the sum of the forecast errors for the gasoline sales time series is 3: thus, the mean or average error is 5/11 = 0.45. Because we do not have sufficient data to produce a naive forecast for the first week of this time series, we must adjust our calcula- tions in both the numerator and denominator accordingly. This is common in forecasting; | ‘we often use k past periods from the time series to produce forecasts. and so we frequently cannot produce forecasts for the frst k periods, In those instances the summation in the nu- ‘erator starts atthe first value of r for which we have produced a forecast (so we begin the ‘summation at = & + 1), and the denominator (which is the number of periods in our time series for which we are able to produce a forecast) will also reflect these circumstances. In the gasoline example, although the time series consists of 12 values, to compute the mean error we divided the sum of the forecast errors by 1] because there are only 11 forecast errors (we cannot generate forecast sales for the first week using this naive forecasting method). ‘Also note that in the gasoline time series, the mean forecast error is positive, which implies thatthe method is generally underforecating: in other words, the observed values tend tobe greater tha the forecasted values. Because positive and negative forecast errors tend to offset one another, the mean error is likely to be small; thus, the mean error is not a 1 very useful measure of forecast accuracy. “The mean absolute error, denoted MAE, is a measure of forecast accuracy that avoids the problem of positive and negative forecast errors offsetting one another. ‘As you might expect given its name, MAE is the average of the absolute values of the forecast errors: FS >t MAE sz) Scanned with CamScanner 746 Chapter 15 Time Series Analysis ond Forecasting This is also refered to as the mean absolute deviation or MAD. Table 15,7 shows that the sum of the absolute valves of the forecast exrony is 41s thus 4 MAE = average of the absolute value of forecast emars = im An Another measure that avoids the problem of pasitive and negative ettors offsetting each other is obtained by compu age of the squared forecast errors. This measure of forecast accuracy, referred to as the mean squared error, is denoted MSE difficult to make comparisons for diferent tne intervals (suchas comming a method of forecasting monthly gasoline sales to a method of foreca y sales) orto make comparisons across different ime series (sich as monthly sales of gasoline and rmonthly sles of ot filters). To make comparisons such as these we need to work With relative or percentage error measures. The miean absolute percentage error, denoted MAPE. is such a measure. To compute MAPE we must frst compute the percentage error foreach forecast: hoo (i For example, the percentage error corresponding to the forecast of 17 in Week 2 is computed by dividing the forecast error in Week 2 by the actual value in Week 2 and multiplying the result by 100. For Week 2 the percentage error is computed as follows: Percentage eror for Week 2 = (2) = (io = 19.05% ‘Thus, the forecast errr for Week 2is 19.05 ofthe observed value in Week 2. A complete summary of the percentage errors is shown in Table 15.7 inthe column labeled Percentage Error In the next column, we show the absolute value of the percentage emor. Finally we find the MAPE, which is calculated as Scanned with CamScanner Try Problem I for practice in computing measures of {forecast accuracy. 15.2. Forecast Accuracy 747 ‘Table 15.7 shows that the sum of the absolute values of the percentage errors is 211.69: thus erage of the absolute value of percentage forecast errors In summary, using the naive (most recent observation) forecasting method, we obtained the following measures of forecast accuracy: MAE = 3.73 MSE = 16.27 MAPE = 19.24% ‘These measures of forecast accuracy simply measure how well the forecasting imethod is able to forecast historical values of the time series. Now, suppose we want to forecast sales for a future time period, such as Week 13. In this ease the forecast for ‘Week 13 is 22, the actual value ofthe time series in Week 12. Is this an accurate estimate of sales for Week 13? Unfortunately there is no way to address the issue of accuracy associated with forecasts for future time periods. However, if we select a forecasting ‘method that works well for the historical data, and we have reason to believe the histo cal patter will continue into the future, we should obtain forecasts that will ultimately be shown to be accurate. Before closing this section, let us Consider another method for forecasting the gasoline sales time series in Table 15.1. Suppose we use the average of all the historical data avail- able as the forecast for the next period. We begin by developing a forecast for Week 2 Since there i only one historical value available prior to Week 2, the forecast for Week 2 js just the time series value in Week 1; thus, the forecast for Week 2 is 17.000 gallons of ‘gasoline. To compute the forecast for Week 3, we take the average of the sales values in Weeks 1 and 2, Thus, n+ ts 19 Similarly, the forecast for Week 4 is 5 42+ he oe Big ‘The forecasts obtained using this method for the gasoline time series are shown in ‘Table 15.8 in the column labeled Forecast. Using the results shown in Table 15.8, we ob- tained the following values of MAE, MSE, and MAPE: MAE 244 MSE =——= 8.10 MAPE = 12.85% Scanned with CamScanner Series Analysis and Forecasting Chapter 1S. Ti c \G FOREC. Y URES OF FORECAST ACCURACY TABLE 15.8 COMPUTING FORECASTS AND MEASU t USING THE DEAL THE HISTORICAL DATA AS THE FORECAST FOR THE NENT PERIOD SY Absolute Absolute Value of Squared piateet “ t Forecast Percentage Percentage Forecast. Forecast Fo cnet Time Series Week Value Forecast Error Error Error 7 21 1700400 4.001600 19.05 19.05 1 3 19 1900 0.00 0.00 0.00 0.00 0.00 ee NE e000 60a 170 21730 2 8 2000-20 200 ant S 16 196-360 3.601295 +2250 an's0 7 2 1900 1.0 1.091.090 5.00 5.00 Ss tetas 635 3 9003003009003 ne 193 06 cnr cae 333 333 2b § .138 440440 1936-2933 BR "pm 30 30 90 he pe Tol 452-2681 89.07 2 M34 We can now compare th accuracy ofthe two forecasting methods we have considered inthis section by comparing the values of MAE, MSE, and MAPE for each, ‘method, Naive Method Average of Past Values MAB 323 244 MSE 1627 8.10 MAPE 1924% 12856 farcath ofthese measures, the average of past values provides mote acute forecasts than using the most recent observation asthe forecast forth next perio In general, if the undying time series is staionary, the average of alte hsricl daa wil provide the ‘mos accurate forecasts Evaluating diferent forecast based on historical accuracy i oly help i his- toveapatems continue in tothe future. AS we noe ia Section ISA, the 12 obser tons of Table 15.1 comprise a stationary ime series. In Section I5'1 we memory that changes in business conditions often result in a time series that isnot stationary, seties and the 10 weeks after signing the new contract, and responding time series pot. Note the change in level in Weel wat Scanned with CamScanner 15.3 Moving Averages and Exponential Smoothing 709 rapidly tothe change in level because it uses only the most recent observation availabe as the forecast Measures of forecast accuracy ae important factors in comparing different forecast ing methods, but we have to be careful to not rely toe ..ily upon them, Good judgment and knowledge about business conditions that might affect the value of the vatlable to be forecast also have to be considered carefully when selecting a method, Historical forecast accuracy is not the sole consideration, especially if the pattem exhibited by the time series js likey to change in the future. Inthe next section we will introduce more sophisticated methods for developing fore casts for a time series that exhibits a horizontal patter. Using the measures of forecast accuracy developed here, we will be able to assess whether such methods provide more ‘accurate forecasts than we obtained using the simple approaches illustrated inthis section. “The methods that we will introduce also have the alvantage that they adapt wel to situa- tions in which the time series changes to anew level. The ability of a forecasting method to adapt quickly to changes in level is an important consideration, especialy in short-term forecasting situations. 15.3 MOVING AVERAGES AND EXPONENTIAL SMOOTHING In this section we diseus three forecasting methods that are appropriate for atime series witha horizontal pattern: moving averages, weighted moving averages, and exponential smoothing. These methods are also capable of adapting well to changes in the level fof a horizontal pattem such as what we saw with the extended gasoline sales time series (Table 152 and Figure 15.2). However, without modification they are not appropriate ‘when considerable trend, cyclical, or seasonal effects are present. Because the objec tive of each of these methods isto “smooth out" random fluctuations in the time series, they are referred to as smoothing methods. These methods are easy to use and generally provide a high level of accuracy fr short-range forecasts, suchas a forecast forthe next time period. Moving Averages “The moving averages method uses the average ofthe most recent k data value inthe time series asthe forecast for the next period. Mathematically, a moving average forecast of order kis as follows: Simos recent data values) _ aif (15.6) forecast of the time series for period + 1 Y,= actual vale ofthe time series in period i wummber of periods of time series data used to generate the Forecast Scanned with CamScanner Con yo now se me averages dec fre Ty Problem 7 Cchoytes 1S Tine Series Analysis ond Ferecostng Tee ttm moving is used becase everytime a new observation Becomes alae fay tims sees, it replaes the oldest observation inthe equation and a new average i gy, puted, Tha, the periods over which the average is calculated change, oF move, With each ensuing peti Toillastrate the moving averages metho, let us return fo the original 12 weeks of gas and Figure 15.1 The time series plot in Figure 15.1 indicates neta of ff cine sates datain Table 15. that the gasoline sales tte series bas a horizontal pattern, Thus the smooth this setion are applicable Tose moving averages to forecast atime series, we must frst select the ender 4, or numberof time seties values to I incline in the moving average, HE only the most re- sent values of the time series are considered relssant, a small value of Kis preferred. Ifa greater number of past values are considered relevant then we generally opt fora larger value of LAs mentioned eatie, atime series witha horizontal pattern can shift to a new level over time. A moving average will adapt to the new level of the series and resume ‘providing goed forecasts in periods. Thus a smaller value of & wil tach shifts im atime ties mone quickly (the nate approach discussed eatlcr 1s actually a moving average for & = 1) Onthe other hand, Large values of & walle more effestive in smoothing out ran «dom fastuations, Thus, managerial judgment hased oan understanding of the behavior cof atime series is helpful in choosing an appropriate valve of d ‘Toollaseate how moving averages can be use to forecast gasoline ales, we will Use a three-week moving average (A = 3). We begin by computing the forecast of sales in Week ‘Fusing the average ofthe time series values in Weeks 1 to 3 4 Rae y= average of Weeks Lip = U7 212! Thus the moving average forecast of sales in Weck 4 is 19 o¢ 19,000 gallons of gaso- line, Because the actual value observed in Week 4 is 23. the forecast ror in Week 4 18 an Bed We neut compute the forecast of sles in Week § by averaging the time series values in Weeks 2-4 +1949 ¥, = average of Weeks 210 as 3 ence the forecast of sales in Week $s 21 and the enroe associate with this forecasts es = 18 31 = ~3. A complete summary ofthe three-week moving average forecasts fo the gasoline sales time series is provided in Table 18.9 Figure 18.7 shows the ori nal time sets plot and the three-week moving average forecasts, Note how the graph fof the moving average forecasts has tended to smooth out the random fluctuations mn the time series “To forecast sales in Week 13. the neut time perio inthe future, we simply compute the average of the time series valoes in Weeks 10,11, and 12 w+Is+2 3 average of Weeks 1010 12 “Tous, the forecast for Week 13 is 19 oF 19.000 gallons of gasoline Scanned with CamScanner 15.3. Moving Averages ond Exponential Smoothing 751 TABLE 15.9 SUMMARY OF THREE-WEEK MOVING AVERAGE CALCULATIONS Absolute Absolute Time Value of Squared Value of Series Forecast Forecast Forecast Percentage Percentage Week Value Forecast Error Error = Error = Error Error 1 2 3 4 4 4 16 1739 1739 5 3 3 9 =1667 16.67 6 -4 4 16 25.00 25.00 7 1 1 1 5.00 5.00 8 0 0 0 0.00 0.00 9 4 4 16 18.18 18.18 10 0 0 0 0.00 0.00 u 5 5 2-33.33 33,33 n 3 3 o 13.64 13.64 0 a 92-2079 = 129.21 FIGURE 15.7 GASOLINE SALES TIME SERIES PLOT AND THREE-WEEK MOVING ‘AERAGE FORECASTS | a 2» i 31s £ ‘Three-week moving g average forces S10 3 a o12345 678 900 23 4 - Scanned with CamScanner 752 CChopler 15. Time Series Analysis ond Forecasting Forecast Accuracy In Section 15.2 we discussed three measures of forecast accuracy mean absolute error (MAE); mean squared error (MSE; and mean absolute percentage error (MAPE). Using the three-week moving average calculations in Table 159, the values ~ for these three measures of forecast accuracy are oa ~ MAPE = 14.36% ln Section 15.2 we showed that using the most recent observation as the forecast forthe In sitarons whereyon "Xt Week (a moving average of order k = 1) resulted in values of MAE = 3.73, MSE = teed compurefnccas. 1627, and MAPE = 19.24%. Thus, in each case the three-week moving average approach ‘ng mets or dfiren has provided more accurate forecasts than simply using the most recent observation as the time pers suchas forecast. Also note how the formulas for the MAE, MSE, and MAPE reflect that our use of elles eenitey | authree-week moving average leaves us with insufficient data to generate forecass for the Foran tert Gist three weeks of ou ime series. Sane oa ena PE. To determine if a moving average with a different order k can provide more accurate ave prefered. forecasts, we recommend using trial and error to determine the value of k that minimizes the MSE. For the gasoline sales time series, itcan be shown that the minimum value of MSE corresponds toa moving average of order k= 6 with MSE = 6.79. If we are wiling to assume that the order of the moving average that i best forthe historical data will also be best for future values ofthe time series, the most accurate moving average forecasts of gasoline sales can be obtained using a moving average of order k = 6, Weighted Moving Averages Amosing average forecast In the moving averages method, each observation in the moving average calculation fender sit aspeciel eves equal weight. One variation, known as weighted moving averages, case ofthe weighted ‘moving averages method inwhich each weghtis qual w I: for example the time period where i, Sinear tend forecast in perio ¢ (ie, the estimated vah by sat (endline + = the slope ofthe linear trendline lue of ¥ in period 1) the ¥-intercept of the - | Scanned with CamScanner 154 ine Trend Projector 759 1 =F eonrespending to the first tne series uJ continues until = n correspond: 10 for the bicycle sales time series) ye series value and In equation (15.10) the tne variable begin obser won (Year | for the bicycle sales time series) to the mont recent time series obvervation (Y ‘Ths, forthe bicycle sales time series ¢ = 1 correspond to the olde = 1 corresponds to the most recent year, Calculus may be used to show that the equa tions given below for by anu by yield the fine that minimizes the MSE, The equations for computing the values of by and by are (5.11) (15.12) where 1 the time period ctual value of the time series in period ¢ ‘n= numberof periods in the time series ¥ = average value of the time series; that is, ¥ mean value of f that Letus calculate by and by fr the bicycle data in Table 15.11; the intermediate surumary calculations necessary for computing the values of by and by are t % Of e 1 216 216 ' 2 ng 458 4 3 255 765 9 4 219 816 16 5 29 1195 % 6 ns 1650 36 1 315 2205 49 8 207 216 6 9 286 214 81 10 314 3140 100 Toul 55 2645 15455 385 Scanned with CamScanner Problem 20 provides addin practice in sing regression ana te estimate the linear trend in atime series daa set. Chopter 15 Tine Series Analysis ond Forecasting And the final calculations of the values of by and by are 58 Pass 0 2645 64S 10 1545.5 ~ (5 SS by = 26.45 ~ 1.10(5.5) = 20.40 Therefore, ee Y= 204+ 1 (15.13) isthe regression equation forthe linear trend component for the bicycle sales time series The slope of I inthis trend equation indicates that over the past 10 years, he frm has experienced an average growth in sales of about 1100 units per year. If we assume thatthe best 10-year tend in sles is a good indicator forthe future, we ean use equation (15.13) to broject the rend component of the time series. For example, substituting ¢ = 11 into equation (15.13) yields next year’s trend projection, f: Fy = 2044 Lad =325 ‘Thus, the linear tend model yields a sales forecast of 32500 bicycles for the next yea. Table 15.12 shows the computation of the minis bicycle sales time series. As previously noted, minimizing sum of squared erors also mini. Imizes the commonly used measure of accuracy, MSE. For the bicycle sales ime series, TABLE 15.12 SUMMARY OF THE LINEAR TREND FORECASTS AND FORECAST ERRORS FOR THE BICYCLE SALES TIME SERIES a Sales (1000s) Forecast Forecast Squared Forecast % y Error Error 216 ws 01 oo1 29 26 03 0.09 255 231 18 304 219 248 29 841 29 259 -20 4.00 25 70 05 05 315 BL 34 1156 207 22 05 025 286 303 -17 289 314 314 00 0.00 Total 30770 Scanned with CamScanner 15.5. Seasonality 761 ‘Note that inthis example we are not using past values of the time series to produce fore~ casts, and sok = 0; thai, we ean produce a forecast for each period of the time series and so do not have to adjust our calculations of the MAE, MSE, or MAPE for &. We can also use the trendline to forecast sales farther into the future. For instance, using equation (15.13), we develop annual forecasts for two and three years into the future as follows: = 24+ 11(12) = 33.6 04 + 1.113) = 347 ‘Note thatthe forecasted value increases by 1100 bicycles in each year. re eee 1. Statistical packages such as Minitab and SAS, 2, While the use ofa linear function to model the as well as Excel, have routines t0 perform trend is common, some time series exhibit a regression analysis. Regression analysis mini- curvilinear (nonlinear) trend. More advanced rmizes the sum of squared error and under cer- texts discuss how to develop nonlinear models tain assumptions it also allows the analyst fo such as quadratic models and exponential mod- ‘make statistical statements about the param- els for these more complex relationships ‘ters and the forecast. ee Inthis section we used simple linear regression to estimate the relationship between the dependent variable (Y, the actual value ofthe time series in period 1) anda single indepen- dent variable (the trend variable), However, some regression models include several inde- pendent variables. When we estimate a inca relationship between the dependent variable and more than one independent variable this is referred to as multiple linear regression, In the next section we will apply multiple linear regression to time series that include seasonal effects and to time series that include both seasonal effects anda linear trend. 15.5 SEASONALITY In this section we show how to develop forecasts for a time series that has a seasonal pattem, To the extent that seasonality exists, we need to incorporate it into our forecasting ‘models to ensure accurate forecasts. We begin the section by considering a seasonal time series with no trend and then discuss how to model seasonality witha Tinear trend. Seasonality Without Trend Letus consider again the data from Table 15.5, the number of umbrellas soldat a clothing store over the past five years. We repeat the data here in Table 15.13, and Figure 15.11 again shows the corresponding time series plot. The time series plot does not indicate any long-term trend in sales. In fact, unless you look carefully at the data, you might conclude that the data follow a horizontal pattem with random fluctuation and that single exponential smoothing could be usd to forecast sales. However, closer inspection of the time series plot reveals a pattem in the fluctuations. Thats, the ist and third quarters have moderate sales, the second quarter the highest sales and the fourth quarter tends to be the lowest quarter in terms of sales volume. Thus, we conclude that a quarterly seasonal patter is present, We can model time series with a seasonal pattem by treating the season as acategor- cal variable. Categorical variables are data used to categorize observations of data. When Scanned with CamScanner rove Orisa sot ee wi A variadle has devels, 1 dum are regina, Soif there ae fo beef siesta Quarter Sales 1 1S 183 106 4 NY 1 us wl 1 4 102 1 138 2 la 3 U3 + w 1 09 2 137 3 135 + 109 1 130 2 16s 3 138 4 SERTES PLOT PAWS IS WT WS 19 0 aT Period a say Variables (sometimes called el variables) Sca8ons, we nod three dummy variables, For instance, in fe quarter to which cach observation gomesponds is teated Ye wh four levels: Quarter 1, Quarter 2, Quarter 3 atl Scanned with CamScanner [Note ha we have mum ered the obsersatons in Table 15.14 periods 110 20, For example, Year 3 Gquarier 3 icobsersation Tl TABLE 15.14 UMBRELLA SAl 155. Seasonality Quarter 4 Ths, tn move the season effects in the unbetla time series we need dummy variables, The thtee domany vartaes can be cored ay follow 1 it period risa Quarter t Quel, 4 [Eitri ris a Quater2 if period is a Quarter 3 otherwise ' Qs, i of sales for period 1, the general form of Ioellas sold to the quarter the sales take place Using to denote the forecasted ¥ the equation relating the number of urn follows, [Note thatthe fourth quarter will be denoted by a seting ofall three dummy variables 00. Table 15.14 shows the umbrella sales time series with the coed values ofthe dummy ‘variables shown, We can use a multiple linear regression model to find the values of bya, and by that minimize the sum of squared erors. For this tegression model Ys the pendent variable andthe quarterly dummy variables Qtr, Qed, and QU3y are the inde- pendent variables. ‘TIME SERIES WITH DUMMY VARIABLES Period Year = Quarter, = QUrl Qu Qu3 Sales 0 125 0 153 106 88 18 161 133 1 138 vy 13 80 109 137 125 109 130 eee oe 128 ceo cos-sse- cco coc eos co -sce-seo-eoo-e e-escHe Scanned with CamScanner "evel ne obina’ the quarterly forecasts for nent year ~ each querer. as shown in (Quarter 3 Quarter 4 105, ss 13 102 13 0 1s 109 ry % QI 5 ing with 2 time series that Sppe0ech will ao work. ‘Sstiss are shown in Table 15.15, ‘nsicates that sles ae lowest inthe seond quarter of Sand Thas we conchae thats easnal parem euits ‘Se time series also as an upward linear trend that will ivecass of quanery sles. Ths is seasonlity withthe Sualiz Seti IST Tre Sm ais pie in eadyer niin See rdevioe et al Hoaen rr to develop ax s arise rprvach for hin Secdon IS foe hendling linear tend, Bee Rawal fs of he regres agate for meting heh the gual seasonal effects and de linear teadin the leven et time acids (15.16) Y= forecast of sees ia periad Qs = Hf peciad cores othe fit quarter ofthe year ccharwise Q6L = Liftine peed scores to che scund guar ofthe year atheise = Hite pend coreg tothe tind quater ofthe year. edhenvise tne pedal Scanned with CamScanner wesfiig ‘SelesTv 765 TABLE 15.15 TELEVISION SET SALES TIME SERIE — Year Quarter Sales (1000s) 1 48 41 60 65 58 52 68 4 60 5.6 Py 18 63 59 80 84 FIGURE 15.12. TELEVISION SET SALES TIME SERIES PLOT Se Set Sates (10008) & U2 Spars G78 Fei ISIS Ion Period — For this regression model Ys the dependent variable and the quarterly dummy variables Qi, Qtr2,, and Qtr3, and the time period # are the independent variables. Table 15.16 shows the revised television set sales series that includes the coded values of the dummy variables and the time period 1. Using the data in Table 15.16 with the regression model that includes both the seasonal and trend components, we obtain the following equation that minimizes our sum of squared errors: 6.07 ~ 1.36 Quel, ~ 2.03 Qur2, ~ 0.304 Qu3, + 0.1460 115.17) Scanned with CamScanner CChopter 15 Time Series Analysis and Forecasting TABLE 15.16 TELEVISION SET SALES TIME SERIES WITH DUMMY VARIABLES AND SS TM PERIOD Period Year Quarter irl Qtr? HES Sales (10045) Problem 28 provides another exanple of sing regression anclss to foes ces ime series da wick ‘ot rend and seasonal fects ' 4 3 6 7 y 9 0 u R 1B im 15 16 1 1 It 0 0 48 2 0 1 0 3 0 0 1 4 0 0 0 2 Il 1 0 0 2 0 1 0 3 0 0 t 4 0 0 0 3 1 1 0 0 2 0 1 0 3 0 0 1 4 0 0 0 4 t 1 0 0 2 0 1 0 3 0 0 l 4 0 0 0 We can now use equation (15.12) to forvast quarterly sles for next year Next year 's ear S forthe television set sues time series; thats, ine periods 17, 18 19 and Forecast for Time Period 17 (Quarter 1 in Year $) Fr = 607 - 1361) ~ 2.030) ~ Forecast for Time Period 18 (Quarter? in Year 5) 6.07 = 1.360) ~ 100) + 0.14617) = 7.19 (1) ~ 0.30440) + 0.14618) = 667 Forecast for Time Period 19 (Quacter3 in Year 5) 6.07 ~ 1.36(0) ~ 2.03(0) — 0.304(1) + 0.146419) = S34 Forecast for Time Petiod 20 (Quarter in Year 3) Fy) = 607 ~ 1.36(0) ~ 2.030) - 0.3040) + 0.14620) = 8.99 ‘Thus, accounting for the seasonal effets and the linear trend in television st sles, the estimates of quarterly sales in Year 5 are 7190, 6670, 8540, and $990, ‘The dummy variables inthe equation atl provide eur equations, oe foreahuaer For instance ifime period comespons to quater Ite esate of quay salesis Quarter I: Sales = 6.07 ~ 1.36(1) ~ 2030) ~ 0.3040) + 0.144 = 4.71 + 0.146 Sima. ite pero ecorespos to quater 2.3.and 4 the estima of que sles ae Quarter2: Sales = 6.07 ~ 1.36(0) ~ 2.03(1) — 0.3040) + 0.1860 = 408 + 0:6 Quate 5: Sales = 607 ~ 1.360) ~ 2.050) ~ 0.3041) + 0.146 = 8.77 + a.tate Quarter 4: Sales = 6.07 = 1 36(0) — 2.030) - 0.3040) + 0.146 = 6.07 + 0. Age The slope ofthe tend line fr each quarry forecast again is 0.46. nating a ‘consistent grovth in sales of about 146 sets per quarter The caly difference in tae four ‘equations is that they have different intercepts, * x Scanned with CamScanner Whenever weategricat rurale such as season nas eels A> Flam smiles are require Summary 167 Models Based on Monthly Data In the preceding television set sales example, we showed how dummy variables can be used to account for the quarterly seasonal effects in the time series. Because there were four levels forthe categorical variable season, hee dummy variables were required. How- ever, many businesses use monthly rather than quarterly forecasts. For monthly data, sea- som isa eaegorial variable with 12 levels, and thus 12 ~ 1 = 11 dummy variables are requited. For example, the HT dummy variables could be coded as follows: Monn = {2 tama 0. otherwise 1 if February a3 set [ei 1 itNovember Monnit = |b fNovember 0 otherwise ‘Other than this change, the approach for handling seasonality remains the same. SUMMARY ‘This chapter provided an introduction to basic methods of time series analysis and forecasting. We first showed tha the underlying pattern inthe time series can often be iden tified by constructing aime series plot. Several types of data patterns canbe distinguished. including a horizontal patter, a trend patter, anda seasonal pater. The forecasting meth- ‘ods we have discussed are based on which of these pattems are present in the time sere ‘We also discussed thatthe accuracy of the method is an important factor in determi ing which forecasting method to use. We considered three measures of forecast accuracy: mean absolute error (MAE), mean squared error (MSE), and mean absolute percentage ‘error (MAPE). Each of these measures is designed to determine how well a particular forecasting method is able to reproduce the time series data that are already available. By selecting the method that is most accurate for the daa already known, we hope to increase the likelihood that we will obtain more accurate forecasts for future time periods. Fora time series with a horizontal patter, we showed how moving averages, weighted moving averages, and exponential smoothing can be used to develop a forecast. The mov- ing averages method consists of computing an average of past data values and then using that average asthe forecast forthe next period. In the weighted moving average and ex ponential smoothing methods, weighted averages of past time series values are used to ‘compute forecasts. These methods also adapt wel o a horizontal pater that shifts to a different level and then resumes a horizontal pattern, FFortime series that have only a long-term linear trend, we showed how regression analy can be used to make trend projections. For a time series with a seasonal pattern, we showed how dummy variables and regression analysis can be used to develop an equation with seasonal effects. We then extended the approach to include situations where the time se ries contains both a seasonal anda linear trend effect by showing how to combine the dummy variable approach for handling seasonality withthe approach for handling a linear trend. Scanned with CamScanner Chops 15 Tine Seis Alsi ond Forecast 78 GLOSSARY ions of data. Used tummy) variable A variable used fo categori7e observations o atime series with a seasonal paler. ate eit if the ime I ove te rend ine lasting more han on Yat / regressionanalysis. Categorical when modeling Cyetca pattern A cyclical sequence of points below and Dependent arable The variable thats bing rfid or explained pe i ihod hat ses a weighted average of past ti Exponential smoothing A forecasting : ots i ened series values asthe forecast; itis a special in which we select only one weight—the weight forth most recent observation. Forecast error The ference between the actual time series value andthe forecast Independent variable A variable used to predict or explain values ofthe dependent vari- able in regression analysis. ‘Mean absolute error (MAE) The average of the absolute value ofthe forecast errors. ‘Mean absolute percentage error (MAPE) The average of the absolute values of the percentage forecast eros. | i series plot shows an alternating ‘Mean squared error (MSE) The average ofthe sum of squared forecast ero. Moving averages A forecasting method that uses the average ofthe most recent data Yaluesin th time series asthe forecast forthe net period Regression analysis procedure foestimatng values ofa dependent variable gventhe values of ne ormore independent variables in manner tat mininies the sum ofthe squared eros. Seasonal pattern A seasonal pattem exists if the time series plot exhibits a repeating pattem over successive periods Smoothing constant parameter ofthe exponential smoothing model that provides the | ‘weight given othe most recent time series value inthe calculation of the forecast value, | Stationary time series A time series whose statistical properties are indepepndent of time. Fora stationary time series, the process generating the daa has a constant mean and the ‘avail ofthe time eis is constant overtime Time series A sequence of observations on avaible measured at sucesive points in time or over successive periods of time. | Time Eel Plot. graphical presentation of tie relationship between time and the time | ‘series variable, Time is shown on the horizontal axis and the ti vr | cieen ime series values are shown | ‘Trend pattern A wend pattem eis if the time series in plot shows gradual shifis - ‘ments to relatively higher or lower values over a longer period et imormere | | | | Weighted moving averages A forecasting i ’ i i§ method that involves selectn Weigh forthe k most recent data values inthe time series and then comping 7 re ‘verge ofthe ofthe values. The sum ofthe weights mustequalone. PROBLEMS 1 Consider the following time series data: i Wek | 100203 4 5g Nae lo Be oa Scanned with CamScanner Bs <= Problems ac 3. SELFIC 4 3. oor Using the naive method (most recent value) as the for ” following measures of forecast accuracy, biielead a, Mean absolute error b. Mean squared error ¢, _Mean absolute percentage error d. What is the forecast for Week 7? Refetothe time series dta in Exercise 1, Using thea forceast forthe next period, compute the eee A eh a ‘a. Mean absolute error Of fonecast acura ‘b. Mean squared error ‘c. Mean absolute percentage error d. Whats the forecast for Week 77 Exercises | and 2 used different forecasting methods, Wi Bae nt for behreldaa Esa mne the teat mek. tng tn ig Consider the following time series data: Month 3 iVainela) ents) 12 119) 23) ie ae Compute MSE wsing the most recent value asthe forecast forthe net psi. We isthe forecast for Month 8? be Compute MSE using the average of al he data available asthe foes fre re period. Whats the forecast for Month 8? e Which method appears to provide the beter forecast? Consider the following ime series data: ‘Week 2 3 4 7 6 Vale | 18 13 16 Construct atime series plot What ype of pattern exists inthe ta? Develop three-week moving average or thi time series. Compe MSE acd a fixe cast for Week 7: User = 0.2 to compite the exponential smoothing valoes forthe time series. Cos- _pute MSE and a forecast for Week 7. dh Compare te three-week moving average forecast with de cexponestial sats forecast using a = 0.2. Which appears to provide Explain, Use tral and eror to find a value of the exponential resulisina smaller MSE than what you calculated for «= Consider the following time series data: smoothing coefficient a tt e 2 12. Month 2 3 5 6 7 Value | 24 mon © 2 2 a Consett series lt What ype of pattem exis inte 3? b ce renner ine os Compe NSE AT cast for Month 8. © cat Ma et exe sing rb forte ie sO pute MSE and a forecast for Month 8. . 4. Compare the three-week moving average forecast arth the exponential ae ing forcast sing «= 0.2. Which appears to provide on MSE? Scanned with CamScanner seies Anais al Fetecasting hoes 1S Tae Se ti to 1 i alo he exponen set A petnaier NSE tan what you cael a . O2, 3. Refer tothe gasoline sale time series data in Table 15.1 IG. ciao yeh al tee tne tas hein ste, WEBEL 1 Gong the MSE tothe futveeh and ive-weck msi . Whaagysrs to be the fest ar of weeks oF asi te, thewmonig average computation? Recall ht MST for the tice W022 8 Referagain tothe palin sales time series data in'Tale 15. FG 8. Cota weight of U2 forthe most receat bssiation, Us or the SELF © at is thd st eet compate a thee-eek time sti, Conte the SE fr the weight moving erage in pa, Bp me weighted moving average to the unviegtted movin MSE forthe unweighted moving average is 10.2 Suppose you ae allowed to choose any weights always tnd ast of weights that woud make the AS average than or an umeighted moving Gesstne Second tos re ighted moving ‘verge fr the } or this werage? Remember that the {ong as they sum to 1 smaller fora average? Why or why not? 9 With asta Kime src dat from Table 15.1, sho the exponent soothing fore. ‘SELFIC ‘casts using a = 0.1, 4 Appling the MSE measure of forecast accuracy, constantofa = O.L oF 1b. Arethe results the same ‘ould you ighled moving woul (0.2 for the gosotine sales time you apply MA Bis used? 10. Withastooting constant ofa = 0.2 equation (15.8) showsth ‘of the gasoline sales data from Table 15,1 js siven by Fy, = 02¥%) 4 O8 Fy, the focast for Week 12is given by fis = 0201+ O8F Ths, could com {wo esuts to show that the forcast for Week 13 canbe witen Mt you prefer a smoothing series? the measure of accuracy? the forecast for Week 13 However, bine these fy ie + 080204 + O8fy) = 0.2%, + OA6Yy + O64, 8. Making use of the fact that fy © 02%10 + OXF (and similarly for Hi and ¥, 19 eas for the 48 moat preceding te storm fr all department tes nthe sth toa sales inthe county forthe four months the Carson Depart Hosed. Carson's managers asked you to anayze these data and deveop ment Store was cl ‘estimates of the lost sales at the Carlson Department ‘Store forthe months of September TABIE 15.18 SALES FOR CARLSON DEPARTMENT STORE (S MILLIONS) Yeard” Year5 Year! = Year2. = Year3 145 231 231 256 180 189 19 28 203 2.02 2a 269 199 223 245 248. 232 239 251 203 220 214 292 237 ' 213 2 2.40 231 | 243 221 250 223 im 190 139 20 | 19 213 229 ea | - 214 256 283 291 December, 4 416 404 435 | SE Ml Mls ee ee Bie canSte Scanned with Camscanner Sehet omer femme Ss Cates Dererment Ste th sommarizes your REED EEG amend Tack Oe folie Sm a eee peter Se fa th tera hance : SEE Soerne teomer sain aes ee SP SGM Ne Sis teete Canker D Tene: ae eae Sater es Dest 5 of cate Sign gue Ist Se hlinens ae om tener ence teeseen mene eae Seep L, Sot Tam Stt2 Faw de Ames geen cies te Tan Amaskaroon Seg 5. We De Tam Aas date tox ame shone Moving Average end = canned wi amscanner Appt 15.1, Forcing wih Excel Dl rc Too 1% ‘ALES DATA IN EXCEL ARRANGED TO USETHE MOVING FIGURE 15.13 GASOLINE AVERAC ICTION TO DEVELOP FORECASTS Bade wom es one et float foods Ou fete Ve QO FD the Moving Average dialog box apes: Enter B2:B13 in the Input Range ‘box Enter 3 in the: Interval box Enter C2 in the: Output Range box. Click OK ripleted this step (as shown in Figure 15.14), C ofthe worksheet asin Figure 15.15. Not that Step 4. When Once you have co the three-week moving average forecasts will appear in column forecasts fr periods of oer lengths can inthe Interval box. ‘be computed easily by entering different valve DIALOGUE BOX FOR A 3-PERIOD MOVING FIGURE 15.14 EXCEL MOVING AVERAGE AVERAGE Scanned with CamScanner Baan f ae Seitna sea) oa li i a rik 2 i ® » 1 is ; 3 a | 5 u » | i i 6 » | I » a U i i 8 ” 1 Exponential Smoothing ‘To show how Excel canbe used for exponential smoothing, we again develop a forecast forthe gasoline sales time series in Table 15.1 and Figure 15.1. We assume that the user ‘has entered the week in rows 2 through 13 of column A and the sales data for the 12 weeks into worksheet ows 2 through [3 of column B (sin Figure 15.13), an tba the smoothing constant is a= 0.2, The following steps canbe used to proce a forecast: Step 1, Select the Data tab Step 2, From the Analysis group select the Data Analysis option Step 3. When the Data Analysis dialog box appears, choose Exponential Smooth ing and click OK ‘Step 4. When the Exponential Smoothing dialog box appears: Enter B2:B13 inthe Input Range box Enter 0:8 in the Damping factor box Enter C2 in the Output Range box Click OK Once you have completed this step (as shown in Figure 15.16, the exponential smooth- ing forecast will appear in column C ofthe worksheet (asin Figure 15.17) Note that the value we entered inthe Damping factor box is 1 ~ a; forecasts for other smoothing Constants can be computed easily by entering a different value for | ~ a inthe Damping factor box. Trend Projection To show how Excel can be used for trend projection, we develop a forecast forthe bicycle sales time series in Table 15.3 and Figure 15.3. We assume that the user has entered the Year (1-10) foreach observation into worksheet rows 2 through 11 of column A and the Scanned with CamScanner jx wih Exel Dats Analysis Tools Appendix 15.1. Forecasting Dots Anais va fAGURE 15.16 EXCEL EXPONENTIAL SMOOTHING DIALOGUE BOX FOR = 020 |ASOLINE SALES DATA AND OUTPUT OF EXPONENTIAL E 15.17 G: ron SMOOTHING FUNCTION IN EXCEL gat ne sales ves into worksheet rows 2 through 11 of colama Bas shown in Figure 15.18. The following steps can be ued to produce a forecast foc Year 11 by tread projection: Step 1. Select the Formulas tab Sep 2, Select two celinthe row where you want he regesion coefficients by and ‘to appear (for this example, choose D1 and El) Step 3. Click on the Insert Funetion key ‘Step 4. When the Insert Function dialog box appears: Choose Statistica inthe Or select a category box ‘Choose LINEST in the Select a function box | Click OK Scanned with CamSéanner Chopte 1S. Tine Series Anais and Forecsing xCE NGE E LINEST FIGURE 15.18. BICYCLESALES DATA IN EXCEL ARRANGED TO USE THE LI FUNCTION TO FIND THE LINEAR TREND gare anes a oon fot pln Fre Oma Fever Wen 9 Qo EB FIGURE 15.19 EXCEL INSERT FUNCTION DIALOGUE BOX FOR FINDING THE TREND LINE USING THE LINEST FUNCTION IN EXCEL See Figure 15.19 for an example ofthis step, ‘Step 5. When the Function Arguments dialog box appears: | Enter B2:B11 in the Known_y's box AZALI! inthe Known 1's box Scanned with CamScanner goes wk Be Oa Anh Fo + Genangnanerse s RaNSarat et heat = het = hae er pouty ges ergs een eed te aan nee Skoveremed veficents by and by in the two cells coun your have pnerated the TEETESSOD i that cell DI contains by and cel EL andispeiat a i ep Its imgoeant 0 8 Be gad he var of Be SO y isin cell D1 andthe vale Se can goerate sat *DI+EL ina bank cell. The forecast for Year Uti this sa which you ete this formu. | ‘de umbrella sakes ime STS Ts cored the yea (1A) foreach ete Se wets Se LN ‘Be gunnery dummy variates ehenns C,D.and Ereqectivel: of codons F. The following steps can projective as shown ia Figere S21, ‘Suep 1 Select the Formas t2> Sot Se pn nannies PS ‘band dt arpear (for this example 3 ‘Step Chek ca the Insert Fenction key i Scanned will cart lanner 76 ‘step 4 When the Insert Fanetion ilo 8 P85 ‘Chose Statistical inthe Or select category bo Choose LINEST in the Select a function box. Click OK Step 5 ‘When the Function Arguments dialog box appears: Emer F:F22 inthe Known_y’s bow Enter C3E22 in the Known x's bot Click OK See Figure 15.22 for an example ofthis sep. ‘Step 6 Hite F2 ky andthe simultaneously hit te Sf, Coetol and Enter keys (Shift + Control + Enter) to create an array that contains the values of the regression coefficient B-Band _At this point you have generated the regression coefficients bs by. Band by in cells Gill selected instep 1 tis important to note that te first cell you select contains by the second cell you selected contains by, the third cell you selected contains y, and the fourth cell you selected contains by (i.e. if you selected cells G1J1 im step 1. the value of ‘by will be in cell G1, the value of by will be in Ht. the vatue of #, will be ia, and the ‘value of by will be in cell JI). Scanned with CamScanner sting with Excel Dota Analysis Tools Appendic 15.1, Fovec fas ta guresnetot mown xs sonst se sala amy seat rete fearsome Toqeenteafoess na blake addtger ent print of snd he product fb and Qtr, and the product ‘of by and Qu, For example, if you wish o use this Tincar trend model to generate a forecast for the first quarter of next year andthe value of by dc ate ae of ysincell HI he valu of sincel ane valu of sin ie enener=[G1+0rH1 +0111 ina blankeel The freas forthe quae fet yea inthis ease 240, wil appear in te blank cel in wich youenter this fonmula Models with Seasonality and Linear Trend used to fit models with seasonality and a linear trend, we evel a forecast forthe umbrella set time srias i Table 15.13 and Figure 15.11. We

    You might also like