Research
Validation: walk-forward backtest
An out-of-sample test of whether historical shape similarity improves directional forecasts beyond a matched comparison in the same market.
Historical similarity is not a prediction or trading signal. These are descriptions of what happened after comparable periods in the past.
No statistically robust advantage was found.
Across 117 tested market-and-window combinations, none survived correction for multiple testing at a 10% false-discovery rate. Shape similarity did not beat the matched comparison in a way that holds up once the number of tests is taken into account.
Technical summary
| Measure | Value |
|---|---|
| Candidate combinations | 153 |
| Tested combinations | 117 |
| Excluded for insufficient out-of-sample data | 36 |
| Positive Brier-skill combinations | 15 |
| Raw p < 0.05 | 1 |
| Discoveries at 10% FDR | 0 |
| Discoveries at 5% FDR | 0 |
| Holm-corrected discoveries | 0 |
| Weighted mean Brier skill | -0.0127 |
Generated 2026-10-05 07:16 UTC · 2026-08-12.1 · raw JSON
How to read this
Every combination of market, pattern length and forward period is a separate test. Running many tests produces some low p-values by chance, so the result that matters is what survives a correction for multiple testing — that is the number reported above.