Explainable Hierarchical Retail Demand Forecasting With Time-Series Foundation Models: Comparison With Statistical And Machine Learning Methods

Abstract

Background

Retail demand forecasting across multiple aggregation levels is essential for consistent strategic and operational decision-making.

Purpose

This study develops an explainable hierarchical forecasting framework for weekly retail sales using Seasonal Naive, XGBoost, and LightGBM base models with Base Forecast, Bottom-Up, MinT-OLS, and MinT-Shrink reconciliation.

Methodology

Unlike single-level retail studies, the framework jointly evaluates accuracy, cross-level coherence, and TreeSHAP interpretability at total, category, and subcategory levels under leakage-safe temporal validation. The Kaggle Superstore Sales Analysis dataset was aggregated into 209 weekly periods containing one total-sales series, three category series, and 17 subcategory series.

Findings

The final four weeks were reserved as a strict holdout, while the preceding 205 weeks were evaluated using five expanding-window rolling-origin folds with a four-week horizon. LightGBM with MinT-OLS was selected by cross-validation, achieving an MAE of 1,344.3577, RMSE of 2,465.7518, WAPE of 48.4886%, MASE of 2.5970, and zero coherence error. When refitted on the full development period, the selected configuration achieved a holdout WAPE of 63.9589%. TreeSHAP identified rolling_mean_13, lag_52, rolling_mean_8, rolling_std_13, and lag_26 as the strongest predictors.

Implications

The dependence plot revealed a nonlinear increase in the contribution of rolling_mean_13, moderated by annual-lag conditions.

Originality

These results demonstrate that reconciliation can improve volume-weighted accuracy and eliminate contradictory forecasts, although intermittent subcategory demand and holdout variability remain important limitations.