Forecasting · 2026
Retail Demand Forecasting
SKU-level demand forecasting with honest evaluation: walk-forward validation across naive seasonal, Prophet, and XGBoost — beating the baseline before claiming anything.
- MAE vs naive baseline (top SKU)
- −74%
- models compared per SKU
- 3
- validation — no leakage
- walk-forward
Architecture
Select a stage to explore it.
01 · Transactions
Real retail transactions cleaned and aggregated to SKU-level demand series — the messy step most forecasting demos skip.
Visuals

01 · Problem
Inventory planning lives and dies on demand forecasts: order too little and you stock out, too much and capital sits on shelves. The question for any forecasting system is whether it actually beats the naive approach a planner would use anyway.
02 · Solution
A forecasting system built around a backtesting framework rather than a single model: naive seasonal, Prophet, and XGBoost compete per SKU under walk-forward validation, and the winner is chosen on out-of-sample error, not preference.
03 · Engineering decisions
The baseline comes first — a model only earns complexity by beating seasonal-naive out of sample. Walk-forward validation over a single train/test split, because retail demand drifts and a one-split score is a lucky draw. And bias is reported alongside MAE, because a symmetric error metric hides the asymmetric costs of over- and under-stocking.
04 · Results
On the highest-volume SKU, Prophet reduced MAE from 702 to 181 against the naive seasonal baseline — a 74% reduction — with XGBoost performing best overall across the evaluation suite. Every number comes from walk-forward backtests, reproducible in CI.
05 · Lessons learned
Forecast accuracy claims are only as good as the validation that produced them. The baseline and the backtest are the product; the model is an implementation detail.