Explainable Machine Learning for Omnichannel Demand Forecasting and Inventory Decision Support in Synthetic Experiments

Abstract Omnichannel demand forecasting should be evaluated not only by predictive accuracy but also by the downstream consequences of forecast-driven decisions. This study evaluates an archived synthetic experimental framework linking one-week-ahead demand forecasting, global model explanation, and a simplified single-period inventory simulation. The accompanying synthetic panel contains 50 stock-keeping units across 10 stores and 104 weekly periods, with 45,000 model-ready observations after temporal feature and target construction. Four forecasting approaches were compared under chronological evaluation: a 13-week seasonal-lag Naïve benchmark, Random Forest, XGBoost, and HistGradientBoosting. Global permutation importance was calculated for XGBoost to examine predictive feature contribution. The archived inventory simulation was treated as exploratory because the stored residual-standard-deviation and safety-stock fields are not fully consistent with the reported forecast-error statistics; these fields are therefore not interpreted as validated safety-stock estimates. HistGradientBoosting achieved the lowest reported MAE (12.136) and RMSE (17.119), while Random Forest produced the lowest archived simulated cost (31.185). Random Forest nevertheless had the highest archived stockout rate (12.185%) and lowest archived fill rate (96.860%). In the XGBoost explanation, the 13-week rolling feature had the largest reported permutation importance (8.121), followed by holiday (2.476) and rolling_4 (1.486). These results show, within the controlled synthetic environment, that predictive accuracy and downstream inventory outcomes can produce different model rankings. The study therefore positions explainability as a decision-support layer rather than evidence of causality or a substitute for planner judgment. External validity is limited because the data are synthetic and the inventory policy is intentionally simplified.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-13
DOI
https://doi.org/10.21203/rs.3.rs-11285040/v1
Primary Topic
Forecasting Techniques and Applications
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Explainable Machine Learning for Omnichannel Demand Forecasting and Inventory Decision Support in Synthetic Experiments

Mohammad Iftekhar Ayub, Paresh Chandra Nath, Md Tarake Siddique, Umama Salsabil Hossain²
3 links
Research Square
Forecasting Techniques and Applications
preprint

Explainable Machine Learning for Omnichannel Demand Forecasting and Inventory Decision Support in Synthetic Experiments

Mohammad Iftekhar Ayub, Paresh Chandra Nath, Md Tarake Siddique, Umama Salsabil Hossain²
preprint en

Abstract

Abstract Omnichannel demand forecasting should be evaluated not only by predictive accuracy but also by the downstream consequences of forecast-driven decisions. This study evaluates an archived synthetic experimental framework linking one-week-ahead demand forecasting, global model explanation, and a simplified single-period inventory simulation. The accompanying synthetic panel contains 50 stock-keeping units across 10 stores and 104 weekly periods, with 45,000 model-ready observations after temporal feature and target construction. Four forecasting approaches were compared under chronological evaluation: a 13-week seasonal-lag Naïve benchmark, Random Forest, XGBoost, and HistGradientBoosting. Global permutation importance was calculated for XGBoost to examine predictive feature contribution. The archived inventory simulation was treated as exploratory because the stored residual-standard-deviation and safety-stock fields are not fully consistent with the reported forecast-error statistics; these fields are therefore not interpreted as validated safety-stock estimates. HistGradientBoosting achieved the lowest reported MAE (12.136) and RMSE (17.119), while Random Forest produced the lowest archived simulated cost (31.185). Random Forest nevertheless had the highest archived stockout rate (12.185%) and lowest archived fill rate (96.860%). In the XGBoost explanation, the 13-week rolling feature had the largest reported permutation importance (8.121), followed by holiday (2.476) and rolling_4 (1.486). These results show, within the controlled synthetic environment, that predictive accuracy and downstream inventory outcomes can produce different model rankings. The study therefore positions explainability as a decision-support layer rather than evidence of causality or a substitute for planner judgment. External validity is limited because the data are synthetic and the inventory policy is intentionally simplified.

Research Square
Washington University of Science and Technology (US), University of the Potomac (US)
Forecasting Techniques and Applications
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon