Skip to content
Open access

A Reproducible, Leakage-Free Pipeline for Censored Demand Forecasting and Inventory Optimization on FreshRetailNet-50K

Aug 2026 · Inventions · 0 citations · 36 references

Abstract

Accurate demand forecasting for perishable goods is complicated by demand censoring: when stockouts occur, observed sales understate true demand and bias both forecasts and replenishment decisions. We present a leakage-free, end-to-end pipeline of three stages: an inverse-Mills-ratio (IMR) recovery heuristic benchmarked against a maximum-likelihood Tobit estimator; a LightGBM ensemble on 117 engineered features with train-only aggregate statistics and recursive multi-step inference; and 20 Newsvendor-family policies, the reported one fixed on a validation fold under a fill-rate floor by a rule encoded in the released code, then evaluated once on a held-out window. On FreshRetailNet-50K (50,000 store-product series, 90 days of daily sales with hourly stock-status vectors), the ensemble attains 33.54% WAPE before recentering; the validation-selected CPU-only calibration improves this to 32.79% and reduces forecast bias from −8.23% to −2.62%. All headline figures come from one fold-causal run: the selected Q90 Direct policy reaches a 0.956 fill rate and 0.241 profit per store-product day, a 46% gain over the static baseline (store-clustered bootstrap 95% CI on the absolute difference, [0.072,0.081]), while meeting the 0.90 service floor. Because evaluation demand is censored observed sales, all profit figures are proxies whose direction of bias relative to latent-demand profit is not identified. Ablations attribute most of the measured gain to the multi-objective ensemble and the service-constrained policy; the full feature set mainly improves bias and stability, and recovery affects bias more than WAPE. The contribution is this leakage-free evaluation protocol and the CPU-feasible operational baseline it supports, released with its leakage tests and clustered-bootstrap inference, rather than a new state-of-the-art WAPE on this benchmark.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.