Post on 08-Jun-2020
transcript
ForecastingInternationalFlowsofReturnableTransportItems
By:PatrickJacobsandRajdeepSingh
Advisor:Dr.EvaPonce
Agenda
2
IndustryOverview
Background
ProjectScope
Methodology
Forecasting
KeyTakeaways
FutureResearchAreas
WhatisaReturnableTransportItem?
3
RTILeasingOverview
4
LogisticsServicesCompany
Customer
TierIICustomer
TierNCustomer
RTILeasingOverview
5
ServiceCenter
Customer1st Tier
Customer2ndTier
ServiceCenter
Customer1st Tier
Customer2ndTier
ServiceCenter
ServiceCenter
FlowServiceCentertoCustomer
CustomertoCustomer
CustomertoServiceCenter
Background
5
Background
6
Background
7
ProjectScope
KeyQuestion• DeterminehowForeignExchangeRateswouldalterdirectionof
ProductFlowsbetweenCanadatotheUS
ProjectObjectives• Develop1-monthaheadforecasttopredictInternationalFlows
betweenCanadaandUSA
• Utilizingmacroeconomicfactorsaspredictivevariables
9
ForecastingwithMacroVariables
VariablesSelectionCriteria
• Relevant• ReadilyAvailable
10
Country Variable Aggregation Level
USA&CA #2 DieselPrices Monthly
USA&CA ForeignExchangeRate Monthly
USA ExportstoCA Monthly
USA ImportsfromCA Monthly
USA GoldPrices Monthly
USA&CA GDP Quarterly
Methodology
• Monthlyvariablelagsrangefrom1to12months• Quarterlyvariablelagsrangefrom3to12months
Results
11 *Correlations where X>.6 or X<-.6
Movement VariableLag
(Months) Correlation
CAtoUSAUSQuarterlyGDP 12 0.7197CanadaQuarterlyGDP 12 0.6868AverageCAtoUSDFEx 5 -0.6995
USAtoCA -- NetInternational --
ForecastingwithMacroVariables
12
ForecastingFramework
13
Identify Accuracy Metrics (MAPE, MAD, MASE)
Identify Relevant Endogenous and Exogenous Independent Variables & Time Lags
Evaluate Stationarity & Level, Trend Seasonality pattern in
Data Set
Subtract the Forecasted US-> CA and CA->US to get ∆ Flows
Choose Top Forecasting Models for ∆ Flows
Develop Univariate & Multivariate Forecasting Methods for US->CA
Approach 2: Best of Breeds
Develop Univariate & Multivariate Forecasting Methods for CA->US
Approach 1: Same Forecasting Method
DataAnalysisLevel,Trend,Seasonality
• Level– meanvalueofY
• Trend– Localmean,periodtoperioddifference
• Seasonality– Repeatingincreaseordecreaseinagiventimeperiod
14
DataAnalysisTrend
Aslightlineartrendexistsintheindividualflows
15
DataAnalysisSeasonality
16
ForecastingModels
• StepwiseRegression• Univariate• Multivariate
• Endogenous• Exogenous(MacroEconomic)
• SARIMASeasonalAutoRegressiveIntegratedMovingAverage
• ExponentialSmoothing- Multiplicative• Standard• StateSpace
17
ForecastingMethodologies
Approach1– UsesamemethodologyforUSAtoCanada&CanadatoUSA
18
𝒀"𝒅𝒆𝒍𝒕𝒂 = 𝒀"𝑼𝑺𝒕𝒐𝑪𝑨 −𝒀"𝑪𝑨𝒕𝒐𝑼𝑺
Seasonal Exponential
Seasonal Exponential
Predicted value of International flow
Approach 2 – Use top performing individual forecasts for USA to ……….. Canada and Canada to US to select “Best of Breed”
𝒀"𝒅𝒆𝒍𝒕𝒂 = 𝒀"𝑼𝑺𝒕𝒐𝑪𝑨 −𝒀"𝑪𝑨𝒕𝒐𝑼𝑺
Seasonal Exponential
Simple Regression
Predicted value of International flow
PerformanceEvaluationMetrics
PerformanceIsMeasuredbyRelativePerformanceonAllThree
19
𝑀𝐴𝑃𝐸 =𝐸𝑡𝐴𝑡
𝑀𝐴𝑆𝐸 =𝐸𝑡
|𝐸𝑡789:;|𝑀𝐴𝐷 = |𝐸𝑡|
Mean Absolute Percent Error
Mean Absolute Scaled Error
Mean Absolute Deviation
Errorinrelationtoactualvalue
ErrorinrelationtoNaiveFt
error
AbsoluteUnitError
PerformanceEvaluationMetrics
Issue:ImperfectMetrics
Solution:CompositeScores• Weighseachmetricevenly&comparesmodels
performancesacrossall3metrics
20
MAPERankMultiplicative: MASERank MADRank
✚
**
Mean: MAPERank MASERank MADRank( )3
✚
ModelSelectionQuantitativeSelection
21
Multiplicative Score RankValidation Rank
Model MAPE MASE MAD MAPE MASE MAD MeanScore Mult ScoreSARIMA(0,1,1)(0,1,0)|SARIMA (0,1,1)(1,1,0) 15.3% 3.84 37340 1 28 1 10.00 28
SeasonalExponential |SimpleRegression 15.8% 3.76 40211 2 26 4 10.67 208
Holt-Winter|SimpleRegression 16.8% 3.63 39365 6 21 2 9.67 252
SeasonalExponential|SeasonalExponential 16.9% 3.27 40314 7 10 5 7.33 350
SeasonalExponential |EndogenousRegression 16.3% 4.09 39895 4 33 3 13.33 396
Validation Rank
Model MAPE MASE MAD MAPE MASE MAD MeanScore Mult ScoreSeasonalExponential|SeasonalExponential 16.9% 3.27 40314 7 10 5 7.33 350
Holt-Winter|SimpleRegression 16.8% 3.63 39365 6 21 2 9.67 252
SARIMA(0,1,1)(0,1,0)|SARIMA (0,1,1)(1,1,0) 15.3% 3.84 37340 1 28 1 10.00 28
SeasonalExponential |SimpleRegression 15.8% 3.76 40211 2 26 4 10.67 208
SimpleRegressionM2Y|SimpleRegression 16.2% 3.69 41462 3 23 6 10.67 414
Mean Score Rank
Differentmodelsselectedwhenusingdifferent
compositescores
22
Seasonal Exponential | Seasonal Exponential was selected due to quantitative and qualitative performance
Model MeanScore MultScore
UpdateRequirement
SoftwareDependency
SeasonalExponential|SeasonalExponential 7.33 350 1 1
Holt-Winter|SimpleRegression 9.67 252 2 1
SARIMA(0,1,1)(0,1,0)|SARIMA (0,1,1)(1,1,0) 10.00 28 1 2
SeasonalExponential |SimpleRegression 10.67 208 2 1
SimpleRegressionM2Y|SimpleRegression 10.67 414 2 1
SeasonalExponential |EndogenousRegression 13.33 396 3 1
SeasonalExponential |SimpleRegressionM3Y 13.67 1015 2 1
ModelSelectionQualitativeSelection
KeyTakeaways
1. Macrovariables*arenoteasilytiedtomicroleveldata2. Methodicalforecastingidentification3. Timehorizonsgreatlyeffecttimeforecastevaluationandperformance
22
Metric SeasonalExponential SARIMA SE PerformanceDifference
MAPE 16.79% 15.3% -8.87%
MASE 3.62 3.83 5.80%
MAD 40492 37340 -7.78%
Metric SeasonalExponential SARIMA SE PerformanceDifference
MAPE 8.23% 4.68% -43.07%
MASE .67 .39 -42.58%
MAD 40492 37340 -43.19%
Aggregation: Monthly
Aggregation: Yearly
BenefitsforReverseLogisticFirms
• IncorporatingSeasonalityinInventoryPlanning
• StrategicPlanningforDemandUncertaintyinReverseLogistics
• MinimizeRTIRepositioningFlowsTransportationCosts
• ImproveBalancingofRTIFlowsacrossNetwork
22
NextStepsUS Origin States - SpreadOrigin US State- Destination Province
DistributionOrigin California – Destination Province
23
FutureResearchAreas1. ForecastingofRTIFlowsatmoreGranularLevel2. MinimizeTransportationCostsbyReducingRTIRepositioning
• PlanFlowstoServiceCenters:StatewideMixforDestinationFlows3. TailorPricingandLeasingContractsusingHistoricalCrossBorderRTIflows
ServiceCenter
Customer1st Tier
Customer2ndTier
ServiceCenter
24
FutureResearchAreas1. ForecastingofRTIFlowsatmoreGranularLevel2. MinimizeTransportationCostsbyReducingRTIRepositioning
• PlanFlowstoServiceCenters:StatewideMixforDestinationFlows3. TailorPricingandLeasingContractsusingHistoricalCrossBorderRTIflows
ServiceCenter
Customer1st Tier
Customer2ndTier
ServiceCenter
24
ReferenceSlides
DataAnalysisForecastability - CV
30
Low Variation Within Sets –High When Combined
Low Variation Within Sets –Consistent When Combined
High Variation Within Sets –High Variation When Combined
CA to US is the easiest flow to predict as the variation is consistent over time
Training Validation Total
UStoCA 12.27% 10.77% 16.17%
CAtoUS 13.17% 12.32% 12.21%
Delta 17.13% 24.58% 19.68%
𝐶𝑉 =𝜎𝜇
EndogenousVariables
• MonthlyNetworkPurchasesofPallets• DomesticMonthlyRTIIssued• DemandGrowth• RTIReturnstoServiceCenters
31
Movement Variable Monthly LAG CorrelationUSA - Domestic Issues 3 0.791CA - Domestic Issues 3 0.783USA Inflows 3 0.771CA Inflows 3 0.766USA Inflows 6 0.785USA - Domestic Issues 6 0.743USA - Domestic Issues 3 0.672CA Inflows 6 0.659
Net International -- --
CA to USA
USA to CA
CAtoUS