All runs

demand-forecast-lgbm-v3

Weekly demand forecast · LightGBM (lags + weather + holidays)

Completed

Metrics vs baselines

MetricModelNaive baselineCurrent methodΔ vs best baseline
WAPE11.8%19.4%16.2%−4.4 pts
MAE (units)8.313.111.0−2.7
Stockout events (held-out)416859−18

Baseline gate

To pass evaluation, the model must beat both the naive baseline and your current method.

Gate passed

Feature importance (SHAP)

lag_1 (last week sales)31.0%
lag_4 (same week, last month)22.0%
temp_c (avg)14.0%
holiday (QC)11.0%
promo_active9.0%
price_index7.0%
site_id6.0%

Model card

Task
Forecasting
Target
units_sold_next_7d per site × SKU
Validation
Time-based split — last 12 weeks held out
Data version
borealdump-2026-09-17
Model
LightGBM (lags + weather + holidays)
Training cost
$4.20 CAD