Backtester

TÓM TẮT

BacktestEngine (core/backtester.py, ~837 dòng) chạy walk-forward validation với configurable windows. Output gồm metrics per fold + breakdowns theo SKU type, event tier, DOW.

Walk-Forward Protocol

Default train windows: [60, 90, 180, 365] ngày
Default test windows:  [14, 30] ngày

Mỗi (train_window × test_window):
  1. Split data: train = data[-train-test:-test], test = data[-test:]
  2. Tính recent_sales_30d per-fold (no leakage)
  3. Train model on train set
  4. Evaluate on test set
  5. Compute breakdowns

Output Metrics Per Fold

Nhóm	Metrics
Primary	wmape, bias, fnr, mae
Event	event_wmape, event_fnr
Breakdowns	sku_type_metrics, event_tier_metrics, dow_metrics, segment_metrics

Tài liệu liên quan

• ML Engine — Tổng quan
• AutoML — Backtest dùng trong improvement loop