Incorporating macro-economic leading indicators in inventory management

Transcription

1 ISIR 16 Presentation: Leading Indicators to Management Incorporating macro-economic leading indicators in inventory management Yves R. Sagaert, Stijn De Vuyst, Nikolaos Kourentzes, El-Houssaine Aghezzaf, Bram Desmet Department of Industrial Management, Ghent University 26/08/2016 ISIR 2016 Ghent University, Lancaster Centre for, Solventure 1

2 When will the next economic crisis hit? Where? For how long? Traditional univariate forecasting techniques do not incorporate context information ISIR 2016 Ghent University, Lancaster Centre for, Solventure 2

3 Research Question Long term sales forecasting are formulated using Historical data patterns (level, trend, seasonality,...) Promotions Judgemental adjustments: Collaborative input from clients Newspapers and industry magazines Rumors in the corridors Judgemental input is known to be biased and inconsistent (Fildes and Goodwin 2007, Trapero et al. 2013) Information of exogenous leading indicators Capturing market sentiment in external big data (Russom et al. 2011) ISIR 2016 Ghent University, Lancaster Centre for, Solventure 3

4 Research Question Can macro-economic indicators improve sales forecasts? What is the real impact on the supply chain inventory? ISIR 2016 Ghent University, Lancaster Centre for, Solventure 4

5 Incorporating leading indicator information Tactical level Plant level Top-down level Evaluation: MAPE and MdAPE ISIR 2016 Ghent University, Lancaster Centre for, Solventure 5

6 Benchmark models Naive model Holt-Winters model Exponential Smoothing LASSO model S P Ŷ i = β 0 + β k D k + β i x ij (1) k=1 j=1 Cost function: n y i β 0 P β p x ip i=1 p=1 p=1 2 P + λ β p (2) ISIR 2016 Ghent University, Lancaster Centre for, Solventure 6

7 LASSO Least Absolute Shrinkage Selection Operator (Tibshirani, 1996) Shrinkage and variable selection Selecting λ through cross-validation Coefficients L1 Norm ISIR 2016 Ghent University, Lancaster Centre for, Solventure 7

8 LASSO Working paper: Sagaert Y. R., Aghezzaf E.H., Kourentzes N. and Desmet B. Tactical sales forecasting using a very large set of macroeconomic indicators. European Journal of Operational Research. MAPE improvement 18.8% on 1-12 months ahead Set of 67,851 indicators Unconditional Final model: indicators selected Employment in automobile National passenger car registrations Consumer Prices Index for solid fuel prices ISIR 2016 Ghent University, Lancaster Centre for, Solventure 8

9 Sales data of 5 plants of a global manufacturer Train period: Test period: Forecast horizon h=1..6 Rolling origin evaluation ISIR 2016 Ghent University, Lancaster Centre for, Solventure 9

10 Empirical results: forecasting accuracy MAPE Naive Holt Winters Exponential Smoothing LASSO Lower level MdAPE MAPE MdAPE Naive Holt-Winters Exponential smoothing LASSO Naive Holt Winters Exponential Smoothing LASSO Forecast Horizon Forecast Horizon ISIR 2016 Ghent University, Lancaster Centre for, Solventure 10

11 Empirical results: forecasting accuracy MAPE Exponential Smoothing LASSO Higher level MdAPE MAPE MdAPE Exponential smoothing LASSO Exponential Smoothing LASSO Forecast Horizon Forecast Horizon ISIR 2016 Ghent University, Lancaster Centre for, Solventure 11

12 Reconsiliation hierarchical forecasting The hierarchy is captured in the summing matrix Reconciliation incorporates 1/MSE of each forecast ISIR 2016 Ghent University, Lancaster Centre for, Solventure 12

13 Empirical results: forecasting accuracy MAPE Exponential Smoothing LASSO Hierarchical LASSO Reconsiled lower level MdAPE MAPE MdAPE Exponential smoothing LASSO Hierarchical LASSO Exponential Smoothing LASSO Hierarchical LASSO Forecast Horizon Forecast Horizon ISIR 2016 Ghent University, Lancaster Centre for, Solventure 13

14 : iterative vs direct forecasting Reformulated LASSO model for each horizon allows for empirical estimation of σ h Direct forecasting: independent across horizons Iterative forecasting: covariances inflate variance ISIR 2016 Ghent University, Lancaster Centre for, Solventure 14

15 simulation Simulation parameters Production standoff t+6 Service level: 0.9, 0.95, 0.99 policy: Make to stock ISIR 2016 Ghent University, Lancaster Centre for, Solventure 15

16 Average inventory per service level ISIR 2016 Ghent University, Lancaster Centre for, Solventure 16

17 LASSO has an improved forecasting accuracy on long-term On short horizons, LASSO leads to service level and inventory improvements ISIR 2016 Ghent University, Lancaster Centre for, Solventure 17

18 Questions? Thank you for your attention! Yves Sagaert - yves.sagaert@ugent.be ISIR 2016 Ghent University, Lancaster Centre for, Solventure 18