Test a strategy against history
Pick a strategy, set your filters and staking, and run the current model over every finalised race in the window. Win bets are selected and settled at the last real bookmaker price captured before the jump — a price you could actually have taken, never the official Starting Price (that's only known after the race is run, so it can't be used without cheating). Place and exotic dividends are estimated unless a real place price was captured too, so treat those as directional. Finalised data runs Fri, 1 Jan – Sun, 20 Sept.
Model predictions are stored for 2024-03-01 – 2026-09-21 — races outside that window (including anything recent) return no bets until the next backfill-predictions run.
⚠ Settled at the official Starting Price — you can bet at SP, but the price isn't fixed until the jump, so it can move a long way from what it was when you'd have actually made the decision to bet. That tends to overstate achievable ROI. Treat results as a rough directional signal, not an achievable-profit estimate.
Max stake caps every bet in dollar terms, however big the bank gets — a real market can't absorb an arbitrarily large bet just because a compounding bankroll has grown large. Without this, Kelly-mode results on a long run aren't achievable in reality.
A backtest is a sanity check, not a promise. It runs the model as it stands today over the past, with no lookahead in the form — but the model is a placeholder, sample sizes get thin once you filter hard, and a strategy that printed last month can bleed the next. How the model works. Gamble responsibly.