Five self-contained studies on 583,774 GB flat racing runners (2016–2026), built around one question: how do you tell a real edge from a leak?
Most published betting analysis reports the winner. This reports the audit — the checks a result has to survive, and the several that plausible-looking findings did not.
I run a live strategy derived from this dataset. Its selection rules are not in this repo. What is here is the methodology that produced it and, more usefully, the methodology that killed most of what came before it.
| Study | Headline | |
|---|---|---|
| 00 | Data integrity | De-vigged market prices give A/E 1.0000 after dead-heat adjustment. If they don't, nothing downstream is trustworthy. |
| 01 | Leak detection | Four kinds of lookahead, each with a test. Conditioning on a post-race column gives A/E 1.96 — a fabricated edge. |
| 02 | RPR circularity | The top-ranked horse on rating-minus-mark wins 98.4% of handicaps. R² 0.86 predicting that quantity from the finishing position alone. |
| 03 | Market efficiency | 15 fundamental features add +0.015 R² over the price. Removing the price costs 0.068. |
| 04 | Run-style NLP | Loose keyword matching is 33.8% precise. A strict clause parse reaches 92.7% coverage and lifts OOS R² from 0.070 to 0.191. |
| 05 | Regime shift | Runners clearing a fixed rating threshold quadrupled in 2022–24 with no change in the sport. |
A performance rating is not a measurement of performance. Racing Post Ratings correlate −0.88 with finishing position and are assigned after the race. The natural reframing — predict how far a horse will exceed its handicap mark, giving a continuous target instead of one bit per race — turns out to be 86% recoverable from finishing position, beaten lengths and field size alone. It is the result restated in pounds, not an easier problem.
The market is very hard to beat, and it is worth quantifying how hard. De-vigged starting prices post AUC 0.780 and move log-loss ~15% off the base rate. The best-rated horse on prior form wins 23.6% of handicaps; the favourite wins 31.7%. Published ratings carry nothing the price has not already absorbed.
Fixed thresholds assume stationarity, and that assumption fails. The gap between ratings and official marks converged by ~2.3 lb from 2022 and partially reverted. Runners clearing any fixed threshold roughly quadrupled. Class mix, field size and turnaround times were flat throughout, and the shift appears within every class — so it is not composition. Notably, the larger movement was in official marks, not the ratings.
Execution assumptions fail as often as modelling ones. Two examples that survived a long way into this project before being caught: a volume-weighted average price used as though it were takeable (it is the average of what others were matched at — you cannot bet at an average), and a maximum-price rule on exchange starting-price bets (the limit is a minimum for backs; no maximum exists). Both produced backtests that could not have been executed.
pip install -r requirements.txt
# place the source files in data/, then:
python src/build_cache.py # ~40s, writes data/built.parquet
python analysis/00_data_integrity.py
python analysis/03_market_efficiency.py # a few minutes, fits GBMsData is not distributed here (Racing Post ratings and Betfair price data are
not mine to redistribute). Expected in data/:
flat_bsp.parquet— runner-level results with Betfair SP, morning WAPflat_2016_2026_csv.gz— source rows; used only for race names, which is the only reliable way to flag handicaps
Every derived feature is rebuilt from source in src/loader.py
so that no number is inherited from an earlier session.
Reusable regardless of sport or market:
- De-vigged prices must give A/E ≈ 1.0 overall. Adjust for dead heats.
- Compute overround on complete books. Summing
1/priceover a filtered subset omits part of the book and understates the margin. - Reconstruct every rolling feature from scratch and require an exact match. Check that first-run rows are empty, and that the feature correlates more with the prior value than with today's.
- Ask of every column: was this knowable before the off? Ratings, in-running comments and finishing positions are not.
- Ask the same of every filter. Filtering on starting price is lookahead; it silently drops drifters, and drifters lose.
- Ask whether the assumed execution price is takeable. Benchmarks and averages are neither takeable nor determined at bet time.
- Split by period, and report the out-of-sample number. Then count how many variants you tried, and discount accordingly.
- Block-bootstrap by day where bets cluster, rather than assuming independence.
- Monitor the selection rate over time, not just the performance of what is selected. A rising rate signals a diluting edge.
- Favourite–longshot bias should be visible. Its absence means the prices and results are not aligned.
Single sport, single market, one dataset. The 2016–2026 window has been used for both development and validation across the wider project, so any performance figure here should be read as an upper bound. Nothing in the methodology is novel to quantitative finance — leak detection, walk-forward validation and block bootstrapping are standard practice. The contribution, such as it is, is applying them to a domain where they are often skipped, and publishing what they killed.