METHODOLOGY

Build long. Test blind. Publish everything.

§1 Data

• Betfair exchange history for GB and Irish racing, April 2015 to January 2026: markets, results, Betfair Starting Prices and traded prices through the day.
• Form: results, going, distance, class, draw, weight, jockey and trainer records.
• Weather at each course: rain in the days before, temperature, wind.

§2 Testing

With hundreds of possible rules, the best-looking one in hindsight is usually luck. So a model is never judged on the data used to build it.

• Walk-forward: to test 2021, the model may only use 2015–2020.
• Holdout: the latest twelve months stay locked away and are used once, at the end.
• Forward test: selections are published with a timestamp before each race; the public ledger is the record.

§3 The pass mark

Set before testing: a positive return after commission in most walk-forward years and in the holdout, on a meaningful sample, with drawdowns and losing runs reported alongside. A model that fails isn't launched.

§4 Costs

Profit and loss is after 5% Betfair commission. Where a strategy relies on an earlier price, we also charge for getting slightly worse prices than the screen showed.

§5 Limits

• Backtests are simulated. Markets change; patterns can stop.
• Losing runs are normal, even with a long-run edge.
• This is statistical analysis, not advice, and guarantees nothing.

§6 Not for

Anyone under 18, anyone who has self-excluded, anyone betting money they can't afford to lose, or anyone looking for income or a way out of money worries. See safer gambling.