71his article considers the supermarket managers problem of forecasting demand for a product as a function of the products attributes and of market control variablesTo forecast sales on the stock ke ID: 897826
Download Pdf The PPT/PDF document "RETAILERPROMOTIONPLANNINGIMPROVINGFORECA..." is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
1 71 RETAILERPROMOTIONPLANNING:IMPROVINGFO
71 RETAILERPROMOTIONPLANNING:IMPROVINGFORECASTACCURACYANDINTERPRETABILITYMICHAEL TRUSOV,ANAND V.BODAPATI,AND LEE G.COOPER his article considers the supermarket managers problem of forecast-ing demand for a product as a function of the products attributes and of mar-ket control variables.To forecast sales on the stock keeping unit (SKU) level,agood model should account for product attributes,historical sales levels,andstore specifics,and to control for marketing mix.One of the challenges here isthat many variables which describe product,store,or promotion conditions ANAND V.BODAPATIis Assistant Professor at the UCLAAnderson School,110 Westwood LEE G.COOPERis Professor Emeritus at the UCLA JOURNAL OF INTERACTIVE MARKETINGJournal of Interactive MarketingDOI:10.1002/dirmost consumer-packaged goods (Srinivasan, Pauwels,sumer behavior, and so on. In the business environ-both retailers and manufacturers; however, theSome recent studies, however, bring to light the topic(e.g., Cooper, Baron, Levy, Swisher, & Gogos, 1999;Divakar, Ratchford, & Shankar, 2005). This interestfor decades, advancement
2 s in technology offeredbetterwaysformana
s in technology offeredbetterwaysformanagerstohandletheplanning process.Topredict demand for future promotion events, theor merchandise category. The output of the data min-pretability of produced knowledge. Specifically, weing algorithm produces an 11.84% improvement overout loss of predictive ability. In our sample, we wereageable and transparent to a market practitioner.recentdevelopmentinthefieldofautomatedpromo- measures: scope and accuracy. Results are discussed,discussed, as are some efficiency issues. We then pre-PROMOTION-EVENT FORECASTINGSYSTEM PROMOCASTconsult the original publication. We offer here a briefgence of scanner technology, businesses have beenCorrective Action Generator. The forecast modulely from a retailers historical records (e.g., unit prices,items manufacturer and merchandise category.RETAILER PROMOTION PLANNINGJournal of Interactive MarketingDOI:10.1002/dir PromoCast Design JOURNAL OF INTERACTIVE MARKETINGJournal of Interactive MarketingDOI:10.1002/dirforecasting model (Cooper & Giuffrida, 2000). As aThe Corrective Action Generator module addresseser overall
3 accuracy.The kernel of the Corrective A
accuracy.The kernel of the Corrective Action Generator moduleis the Knowledge Discovery using SQL(KDS) data(explained in Technical Appendix,Part A). The KDS algorithm is used in the PromoCastVery High ANDGeneral Foods ANDCulver City, #231UNDER_4_11 OVER_4_11 DCS stands for the triple Department-Commodity-Subcommodity, identifying a particu-TPR identifies the level of the Temporary Priceprice reduction. Values for this variable have beenNone, Low, Medium, High, and Very High.Mfr identifies the manufacturer of the givenStore Node allows for store-specific effects.(OVER_4_11 class).keting applications to date. Algorithm application toEnterpriseMiner, SOMine, CN2, Ripper, Apriory, Technical appendices are available from the authors upon request. Corrective Action Generator strongly contributes toforecast accuracy, the module has certain limitationsSEARCHING FOR RAREPATTERNSexplain specific outcomes in residuals. Specifically, inthe example presented earlier, the discovered patternVery High, Mfr Culver City, #231}.tial tasks in data mining. To ensure generation of
4 theimum support (Han, Wang, Lu, & Tzvetk
theimum support (Han, Wang, Lu, & Tzvetkov, 2002).users goal. The higher the value, the fewer rules arelem of defining minimum support is quite subtle: Aand outcome (rules) interpretability. The latter couldWestartwiththehypotheticalexampleofinventoryplanningforanupcomingpromotioneventintheicecreamproductcategory.Letusconsiderasupermar-ketwhichcarries40differentflavorsofBen&Jerrysicecream.Thestoredatabasecontainshis-toricalrecordsonthepastpromotionsandinventorystatusforall40flavors.Ithappensthatmostflavorsexceptcherryandcoffeeareunderforecastedbythestatisticalmodelandwouldresultinoutofstockbeforethepromotionends.Assumethatthestoredatabasecontainsrecordsfor200promotioneventswherenoneofanyparticularflavorofBen&Jerrysicecreamappearsmorethan30times.Finally,letsassumethattheminingapplicationusesaminimumsupportlevelof50occurrences.ThentheresultingruleforBen&Jerrysicecreammaylooklike:ICE CREAM ANDBEN & JERRYSCREAM, MFR = BEN & JERRYS} appears in asso-ICE CREAM ANDBEN & JERRYS ANDFLAVOR Note, however, that the following set of rules wouldproduce more accurate forecast for B
5 en & Jerrys iceICE CREAM ANDBEN & JE
en & Jerrys iceICE CREAM ANDBEN & JERRYSFLAVOR ALLBUT CHERRY AND COFFEERETAILER PROMOTION PLANNINGJournal of Interactive MarketingDOI:10.1002/dir JOURNAL OF INTERACTIVE MARKETINGJournal of Interactive MarketingDOI:10.1002/dirICE CREAM ANDBEN & JERRYS ANDFLAVOR CHERRY OR FLAVOR BEN & JERRYS, FLAVOR CHERRY} and {DCS BEN & JERRYS, FLAVOR with an OR-type construct; however, there are cer-tain considerations to keep in mind. Ahistorical pro-importance. To provide such scalability and also takewould be dropped by the Corrective Action Generatorcream (e.g., {CHERRY OR COFFEE}, {CHERRYOR COFFEE OR ORANGE}, {STRAWBERRYOR VANILLA OR ORANGE}, etc.), we first pro-SUBCOMMODITY, MANUFACTURER, {SET OFFLAVORS}{SETOFSUBCOMMODITIES},MANUFACTURER,FLAVORSUBCOMMODITY, {SET OF MANUFACTU-RERS}, FLAVORappealing. Indeed, it is consumerstaste preferencesIn application to the PromoCast Corrective ActionGenerator, grouping helps in capturing rare geo-Totest the performance of the proposed approach, weneed to establish some benchmarks. We propose thatSecond, use the Co
6 rrective Action Generator to minethe r
rrective Action Generator to minethe recommendations. Finally, repeat the described rithm scalability. Weuse the terms OR, grouped, and composite interchange- Forecast module. Atotal of 4,195 records were3,184recordsintheholdoutsamplerequiredcorrectiveRecall that the Corrective Action Generator makesto the mining module (and, accordingly, does notThe Corrective Action Generator first identifiesthen applies adjustment according to these rules. WeAccordingly, the percentage improvement in the fore-RESULTSThe original Corrective Action Generator attempted(11.84%). The Corrective Action Generator created aTheextendedversionoftheKDSalgorithmproducedadditional85groupingrules,foratotalof156rules.TheCorrectiveActionGeneratorwithgroupingsupportattemptedtocorrect2,805of4,000recordsintheholdoutsample.Basedonknowledgecontainedincom-positerules,2,581adjustmentsweremade;therest224predictionscamefromsimplerules.Notethatintheoriginalversionofthealgorithm,allpredictionsarebasedonsimplerules.Thesuccessratewas78.29%.Overallimprovementintermsofrecordswas1,592of3,184(50.0%).Whencomparedwithres
7 ultsproducedbytheoriginalCorrectiveActio
ultsproducedbytheoriginalCorrectiveActionGenerator,weobservesignificantimprovementsinperformance.RULES SPACE OPTIMIZATIONIntheprevioussection,weshowedhowminingandgroupingofrarepatternssignificantlyimprovesforecastperformance.Inthissection,webrieflytouchonsomeissuesthatarecommontosystemsbuiltonrule-inductionalgorithms.Asdescribedearlier,rule-inductionalgorithmsproduceknowledgebymeansofscanningthroughhistoricalrecordsinsearchforpat-terns.Knowledgeisformalizedintheformofrulesinthefollowingformat:.Inatypi-calreal-lifeapplication,itiscommontohavealargenumberofrulesgeneratedbypopulardataminingalgorithms(includingtheonedescribedinthisstudy).Largenumbersofrulestendtohavehighdegreesofredundancy.Byredundancy,wemeanthatmultiplerulesapplytothesamerecordandproduceidenticalpredictions.Toclarifythisidea,wesuggestthefollowinganalogy:thenumber of rules constituting knowledge might beRETAILER PROMOTION PLANNINGJournal of Interactive MarketingDOI:10.1002/dir JOURNAL OF INTERACTIVE MARKETINGJournal of Interactive MarketingDOI:10.1002/dirin the latter case, it may directly impact a systemsoverall usability
8 . Many business settings dictatetheneces
. Many business settings dictatethenecessity for a human expert to conduct therulesscreening. Mindless or blind application oftomer, such as dynamic pricing or product recommen-Wesuggest that by applying simple optimizationnificant loss of predictive ability. The resultinglyze and to perform screening. We demonstrate thisvalue in terms of prediction accuracy. The situationeachvalue of Wedo an almost identical process here, except thatidentifying the best subset of rules. We choose variousapplicable only to very few records, however, is notBEN & JERRYS,FLAVOR CHERRY}generated rules shown in Table 1.RULE IDDCSMFRFLAVORPREDICTION1IcecreamBEN & JERRYSCherryUnder 10 Cases2IcecreamANYCherryUnder 5 Cases3IcecreamBEN & JERRYSAnyUnder 5 Cases4AnyBEN & JERRYSAnyOver 3 Cases5AnyHaagen-DazsCherryUnder 5 Cases6AnyHaagen-DazsCherryOK7YogurtAnyCherryOver 5 Cases 8AnyAnyVanillaOver 10 Cases TABLE 1 ItiseasytoseethatRules1through4canbeappliedtotherecordofinterestwhileRules5through8arenotapplicable,astheydonotmatchapatterninthefocalrecord.NotethatRule2,forexample,canbeappliedtoanymanufacturerwhopro
9 ducescherryicecream,andRule3isapplicable
ducescherryicecream,andRule3isapplicabletoallflavorsofBEN&JERRYSicecream.Themea-sureofthescopetellsushowmanyrecordsinthecalibrationdatasettowhichtheparticularrulecanbeapplied.Consequently,thescopeofarulessetisdeterminedasthecombinedscopeofeachindividualruleintheset.Tofind the best subset of rules, we follow the tradi-managerial judgment. To minimize generated by the Corrective Action Generator.ErrorrateofthebestRuleSetwithKrulessignificantly lower number of rules. To test this(the optimal size of rules set). As10 is 19 rules. All runs were per-Results are reported in Table 2.terms of accuracy (Actually, we even gained tworecords.) At the same time, we were able to shrinkWesummarize the optimization results in Figure 3 byplottingthenumberofcorrectpredictionsasafunctionRETAILER PROMOTION PLANNINGJournal of Interactive MarketingDOI:10.1002/dir behavior. 1163146617691106121136151 Rules Space Optimization JOURNAL OF INTERACTIVE MARKETINGJournal of Interactive MarketingDOI:10.1002/dirpossibility of human error. The most common problemis the experts overconfidence (for some examples inVallej
10 o, & Barlas, 1999). Both types of rules
o, & Barlas, 1999). Both types of rulesthe onesdriven by data from the calibration sample. However,in the case of human generated rules, it is the expertsthe example mentioned earlier, as well as rules mis-CONCLUSIONthat they are best suited for. Traditional market-Berry, M.J.A., & Linoff, G.S. (2004). Data MiningTechniques: For Marketing, Sales, and CustomerWiley.Cohen, W.W. (1996). Learning Trees and Rules With Set-Valued Features. Proceedings of the 8th annual confer-ence on Innovative Applications of Artificial Intelligence,Cooper, L.G., Baron, P., Levy, W., Swisher, M., & Gogos,P.(1999). PromoCast: ANew Forecasting MethodforPromotion Planning. Marketing Science, 18(3),oftherulessetsize.ThegraphsinFigure3andTable2we are able to achieve prediction results very closetooneproducedbyafull(i.e., 156 rules) rules set. TABLE 2Hold-Out Sample Performance WithDifferent Penalty WeightsOPTIMALACCURATERULES SET SIZEPREDICTIONS352,198332,198212,199202,192172,192142,17952,128 42,128Number of Correct Predictions as a Function of the Rules Set Size 1265176101126151 Cooper, L.G., & Giuffrida, G. (2
11 000). Turning DataminingInto a Managemen
000). Turning DataminingInto a Management Science Tool. Management Science,Divakar, S., Ratchford, B.T., & Shankar, V. (2005).CHAN4CAST: Amultichannel, multiregion sales fore-Giuffrida, G., Chu, W., & Hanssens, D.M. (2000). MiningClassification Rules From Datasets With Large Numberof Many-Valued Attributes. In Lecture Notes inComputer Science (Vol. 1777, pp. 335 349). New York:Han, J., Wang, J., Lu, Y., & Tzvetkov, P. (2002). Mining Top-K Frequent Closed Patterns Without Minimum Support.Forecasting, June July, (17), 327 346.Klayman, J., Soll, J.B., González-Vallejo, & Barlas, S.(1999). Overconfidence: It Depends on How, What, andWhom You Ask. Organizational Behavior and HumanKrycha, K.A. (1999). Übung aus Verfahren dererfahren derEffects of Promotional Activities on Sales]. UniversitätWien, Institute für Betriebswirtschaftslehre.Meehl, P.E. (1954). Clinical Versus Statistical Prediction.Srinivasan,S.,Pauwels,K.,Hanssens,D.M.,&Dekimpe,M.(2004).DoPromotionsBenefitManufacturers,RetailersorBoth?ManagementScience,50(5),617 629.RETAILER PROMOTION PLANNINGJournal of Interactive MarketingDOI:10.1