Validation
Walk-forward backtesting — folds before faith
A model that only looks good in a single in-sample window is a story, not a desk tool. Use rolling train/test folds, then pay costs and embargo leakage.
Example fold table
| Fold | Train | Test | Embargo | Note |
|---|---|---|---|---|
| F1 | 2018–2020 | 2021 H1 | 5d | Baseline regime mix |
| F2 | 2019–2021 | 2022 H1 | 5d | Inflation shock year |
| F3 | 2020–2022 | 2023 H1 | 5d | Post-squeeze vol |
| F4 | 2021–2023 | 2024 H1 | 5d | Narrow-leadership tape |
| F5 | 2022–2024 | 2025 H1 | 5d | Hold-out stress |
Leakage & cost checklist
- No future bars in features or labels
- Purged gaps between train and test
- Corporate actions / splits handled consistently
- Commission + spread + slippage deducted
- Report live degradation expectation (often −8 to −12 pts)