TahaImran1/ai-demand-forecasting-weights

AI Demand Forecasting: Trained Model Weights

This repository hosts the audited, production-grade model checkpoints and fitted preprocessing scalers for the AI Demand Forecasting & Inventory Decision Support project (Deliverable 2).

All models are trained on real Kaggle benchmark datasets under strict chronological validation splits (zero lookahead leakage).

Directory Structure

models/
├── a_lgbm_p10.txt                          # Pipeline A (Rossmann) LightGBM P10 Regressor
├── a_lgbm_p50.txt                          # Pipeline A (Rossmann) LightGBM P50 Regressor
├── a_lgbm_p90.txt                          # Pipeline A (Rossmann) LightGBM P90 Regressor
├── a_mlp.pt                                # Pipeline A (Rossmann) PyTorch QuantileMLP Checkpoint
├── b_lgbm_p10.txt                          # Pipeline B (Store Item) LightGBM P10 Regressor
├── b_lgbm_p50.txt                          # Pipeline B (Store Item) LightGBM P50 Regressor
├── b_lgbm_p90.txt                          # Pipeline B (Store Item) LightGBM P90 Regressor
├── b_mlp.pt                                # Pipeline B (Store Item) PyTorch QuantileMLP Checkpoint
└── scalers/                                # Leakage-free continuous feature scalers (train split only)
    ├── rossmann/standard_scaler.pkl
    ├── rossmann_standard_scaler.pkl
    ├── store_item/standard_scaler.pkl
    └── store_item_standard_scaler.pkl

Evaluated Test Performance Summary

• Pipeline A (Rossmann Store Sales, 1,115 stores, 46,830 holdout instances):
  • LightGBM: 8.67% WMAPE, 80.31% Interval Coverage $[P_{10}, P_{90}]$, 98.04% Stockout Recall (AUC-ROC 0.9875).
  • QuantileMLP: 12.59% WMAPE, 99.89% Coverage, 100.00% Stockout Recall.
• Pipeline B (Store Item Demand, 500 series, 45,000 holdout instances):
  • LightGBM: 10.90% WMAPE, 78.11% Interval Coverage $[P_{10}, P_{90}]$, 98.62% Stockout Recall (AUC-ROC 0.9877).
  • QuantileMLP: 13.05% WMAPE, 95.46% Coverage, 99.90% Stockout Recall.