Uber动态溢价分析报告

PeekingBeneaththeHoodofUberLeChen,AlanMislove,ChristoWilsonNortheasternUniversityPublishedatACMIMC2015InTokyoTheresearchwasledbyChristoWilsonLeChen北京邮电大学信息工程东北大学博士计算机科学现任Facebook研究员主要研究：建立方法提高黑箱算法的透明性和公平性ChristoWilson东北大学教授计算机系主要研究：大数据、安全、隐私的交叉领域AlanMislove东北大学副教授计算机科学主要研究：网络服务的安全性和隐私性WhatisUber?•Selectapickuplocation•ChoosetypeofcarUberX:cheapsedanUberSUV:cheapSUVorvanUberBlack:fancysedanUberXL:fancySUVUberPool:carpoolwithsomerandosEtc…•RequestandrideawaySimpleandConvenient,Except…MarketplacesTransparent•Supplierssettheirownprices•CustomersmayobserveallproductsandtheirpricesOpaque•Supplyanddemandarehiddenfromcustomers•Suppliersdonotchoosetheirownprices•PricesaresetbyanalgorithmGoals•DeterminehowthesurgepricingalgorithmworksDoesitworkthewayUberclaims?Isitresponsivetochangesinsupplyanddemand?•Cansurgesbepredictedand/oravoided?•ImpactofsurgesondriversandpassengersDataCollectionMeasuringSurgesAvoidingSurgesImpactofSurgesConclusionsDataCollection•Uber’sofficialsurgepricingpatentsaysthecalculationisbasedonsupply,demand,andotherfactors•Howcanwecollectthisdata?1.UberAPIPros:easytouse,includessurgemultipliersandEstimatedWaitTimes(EWT)Cons:nocars,demand,orsupplyinformation2.UberRiderAppUberApp•PingsUber’sserversevery5seconds•8nearestcars•EstimatedWaitTime(EWT)•SurgemultiplierLimitations•MeasuringsupplyisstraightforwardSupply=observedcarsontheroad•MeasuringdemandistrickyCarsmaygetbooked…Orjustgooffline…Ordriveoutofthearea•WecanonlyestimatedemandFulfilleddemand=numberofcarsthatgoofflineUpperboundontruefulfilleddemandLimitedVisibility3pmonSundayRecall:theUberapponlyseesthe8closestcarsLimitedVisibilityr’r5pmonMondayHowfarapartshouldweplaceourmeasurementspoints?RadiusMeasurementsCarsobservedbyallapps:RadiusMeasurementsValidation•Validatedmethodologyusingground-truthdatafromNYCtaxisBuiltan“Ubersimulator”andusedourmethodstomeasurethetaxisObserved97%ofsupplyand95%ofdemandFinalDataCollection•CollectedfourmonthsofdatafromMidtownManhattanandSanFrancisco2monthsfromeachcity,43measurementpoints2ndand3rdlargestUbermarketsVerydifferentpublictransportoptions•Radiusexperiments247metersinMidtownManhattan387metersindowntownSanFranciscoExampleMeasurementGrid•43measurementpoints•Collected2monthsofdatainMidtownEthics•WedidnotcollectanypersonalinformationaboutUberdriversorpassengers•Weneverbookedanyrides•WedidnotinduceanysurgesWeplaced40“users”inrandomlocationswithnosurge,inthemiddleofthenightDidnotobservesurgesforonehourRepeated100timesatdifferentlocationsandhoursDataCollectionMeasuringSurgesAvoidingSurgesImpactofSurgesConclusionsResearchQuestions•Howmuchandhowoftendoesitsurge?•Howlongdosurgeslast?•Howdosurgepricesvarybylocation?•WhatfeaturesdoesUberusetocalculatesurgeprices?CollectedDataHowmuchdoesitsurge?•14%ofthetimeitissurginginManhattan•57%ofthetimeinSF•SurgemultiplierstendtobehigherinSFHowlongdosurgeslast?•Noisiness:70%ofsurgeslast=10minutes•Staircase:surgeslastmultiplesof5minutesHowdoessurgevarybylocation?MatchNoMatchMatchHowissurgecalculated?•ManypossiblevariablesSupply,demand,EWT,etc.•Usecross-correlationtoperformtime-seriesanalysisTime12pm12:0512:1012:1512:2012:2512:30SurgeSupplyHighcorrelationat-10minutesCross-correlation(Supply–Demand)vs.SurgeEWTvs.Surge•ModeratelystrongcorrelationswhentimedifferenceiszeroSuggestsUberusesdatafromthelast5minutewindowwhencalculatingsurge•ZerocorrelationinothertimewindowsSurgepricingalgorithmisresponsivebutnoisyDataCollectionMeasuringSurgesAvoidingSurgesImpactofSurgesConclusionsCanwepredictsurges?•UsefulvariablesSupply,demand,EWTPrevioussurgemultiplier(s)•PredictivemodelsLinearandnon-linearregressions•PerformanceR2rangesfrom0.37–0.57Theseresultsareterrible,i.e.wecannotreliablypredictsurges•Missingsomekeyvariable(s)Unfulfilleddemand:howmanypeopletriedtobookaridebutcouldn’t?AvoidingSurgesNoSurge!EWT=6minNoSurge!EWT=3minAvoidingSurges•10-15%chanceyou’llsavemoney•Savingsupto50%byavoidingsurgesDataCollectionMeasuringSurgesAvoidingSurgesImpactofSurgesConclusionsImpactonSupplyandDemand•WhydidUberimplementsurgepricing?ToequalizesupplyanddemandReducedemandbyraisingpricesIncreasesupplybyincentivizingdrivers•Ineconomicterms,surgepricingisaboutincentives•Aretheincentivesprovidedbythesurgepricingsystemworking?StateTransitions•Ifoneareaissurging,weexpectthefollowingfivethingstohappen1.(Supply)New:carsshouldprefertocomeonlineinthesurgingarea2.(Supply)Move-in:carsshoulddriveintothesurgingarea3.(Supply)Move-out:fewcarsshoulddriveoutofthesurgingarea4.(Demand)Booked:fewercarsshouldgetbookedinthesurgingarea5.(Demand)Old:morecarsthatbeganinthesurgeareashouldremainafter5minutesComparingStateTransitions1.(Supply)New:carsshouldprefertocomeonlineinthesurgingarea2.(Supply)Move-in:carsshoulddriveintothesurgingarea3.(Supply)Move-out:fewcarsshoulddriveoutofthesurgingarea4.(Demand)Booked:fewercarsshouldgetbookedinthesurgingarea5.(Demand)Old:morecarsthatbeganinthesurgeareashouldremainafter5minutesStateChangeWhenAreaisSurgingExpected?New+2%YesMove-in-13%NoMove-out+14%NoBooked-7%YesOld+14%