Walk-forward backtest
Method
- 1. Selections are recorded before results are consulted. The analyser reads no finishing position, margin, dividend or starting price for the race it is grading, and every historical aggregate it builds is bounded to dates strictly before that race.
- 2. Periods are chronological. Days are split 60 / 20 / 20 in time order. Randomly mixing races would let a model learn from the future of a horse it is about to assess.
- 3. Weights may be tuned on Train only. Validation selects between candidates. Test is opened once, at the end, and never used to choose anything.
- 4. Improvement must survive the interval test. A higher capture rate whose 95% confidence interval overlaps the baseline is not an improvement — it is noise wearing a better number.
- 5. The production model stays until beaten. No weight is replaced on the strength of training-period gains alone.
For archived races the captured price may be the starting price rather than the price showing when the analysis would have run. SP forms before the off, so this is not result leakage — but it is a sharper signal than a morning-of assessment would have had, and it will flatter any market-derived feature. Treat market-driven gains with suspicion until the same evaluation has been run with price excluded.
Not enough graded racing to report on yet.
A walk-forward split needs at least 30 distinct graded days before the periods mean anything, and several hundred races per period before the intervals are tight enough to act on. The crawlers fill this automatically — this page will populate itself rather than show a figure it cannot stand behind.
If this has been empty for more than a day, check the crawler status on the analysis page. A stalled crawler shows in red there.