<![CDATA[Retail sales are a censored view of demand when products become unavailable. This study evaluates profile-weighted temporal reconstruction and relational forecasting on FreshRetailNet-50K, then examines cross-dataset error tradeoffs on M5. FreshRetailNet contains 50,000 store–product series, 4.85 million daily records, hourly sales, and stockout labels. Category-hour profiles estimate missing intraday shares, while a trailing 28-row own-series prior supplies the demand scale. Artificial contiguous-hour masking on days without labeled open-hour stockouts yields reconstruction WAPE of 8.82%–21.32%; it does not validate natural-stockout demand recovery. On 206,562 observable rows in the seven-day official test, calibrated relational forecasting achieves RMSE 0.6761, MAE 0.3621, and WAPE 29.27%, versus 30.54% for observed-sales boosting. Recovery alone does not improve WAPE, and the incremental graph contribution is small. Two earlier seven-day windows support the direction of the calibrated comparison but do not assess seasonal stability. M5 evaluation covers 30,490 series and 28 rolling one-day-ahead origins. Fixed-budget 1:1 pooling lowers relational-model RMSE from 2.1075 to 2.0716, while increasing WAPE from 71.67% to 72.21%. Alternative sampling ratios and full-target augmentation show the same metric tradeoff. Dense series receive the strongest squared-error benefit, without a WAPE advantage. The findings distinguish controlled reconstruction accuracy from observable-sales prediction and show that cross-retailer pooling does not uniformly improve forecasting.]]>
Copyrights © 2025
