Horse racing as an ML ranking problem: 1.18M runners, walk-forward validation and a very strong market baseline [D]

Hi 👋 long time lurker

It’s been a few years since the last post about Horse Racing within this sub.

A few years ago I was reading a popular science book and came across the story Bill Benter and the statistical models his team developed for Hong Kong racing.

What stayed with me wasn’t simply the idea of using a model to gamble. Horse racing seemed like an unusually difficult applied ML problem: variable-sized fields, one winner per race, highly correlated competitors, missing and changing data, human decision-making, non-stationarity and an extremely efficient market baseline. What’s more, Hong Kong racing is a much smaller domain - two tracks, whereas in the UK & Ireland (where I am based and where my interest lies) there are over 80 tracks with a total track/distance/race type combinations exceeding over 900 configurations.

This ML problem became a personal project of mine called Hoofs, focused specifically on British and Irish racing.

I’ve assembled around 1.18 million historical runner records covering approximately ten years of racing. From this I’ve built a unified feature bank containing roughly 1,700 potential signals per runner, although individual models use much smaller selected subsets and not every newer data source is in the production models yet. I do have fairly novel data which I recently acquired and the feature engineering is what I’m going to be working on next.

The primary models estimate win and place probabilities at runner level, after which the runners are ranked within each race. I also have a separate race-level confidence model based on features such as field size, probability concentration, entropy and the separation between the leading runners.

The public Top 1–3 rankings are deliberately market-agnostic. Market information is evaluated separately as a benchmark and in experimental late-market models.

Everything is evaluated chronologically. Each walk-forward fold is trained only on earlier seasons, with out-of-fold calibration and explicit checks to prevent future information entering historical features.

I track AUC, log loss, Brier score and ranking metrics, but I also score the outputs in racing terms: how often the winner appears at rank 1, within the top three and within the top five.

On one large 2018–2025 benchmark covering approximately 886,000 runners and 94,000 races:

* Model-only win AUC: approximately 0.729

* Model-only place AUC: approximately 0.708

* Market-only win AUC: approximately 0.790

* Market-only place AUC: approximately 0.762

The strength of the market baseline has probably been the most difficult part of the project. Producing a model with decent discrimination is one thing; extracting information not already reflected in the price is quite another. Typically though I find positive EV before the market has fully formed.

I have been publishing daily reports for free, since the start of the year. And the whole project has developed quite a lot since then, introducing new models and different tankers etc and it’s constantly evolving.

I began noticing some degradation in the live strike rates. Going back through the pipeline exposed gaps and inconsistencies in parts of the historical data and feature coverage.

That led to a much larger rebuild: consolidating the raw datasets, rebuilding the feature bank, tightening the chronological lineage and retraining the model families.

The rebuilt reports had their first live day today. The Top 1 selection won 10 of 23 races after one non-runner, giving a 43.5% strike rate, while the winner appeared somewhere in the Top 1–3 in 16 of 24 races.

Practically speaking, I use the daily reports as first layer of analysis and I use market data as the second layer to dictate actual betting.

One reason I wanted to post this is that, I was surprised by how little discussion I could find about horse-racing models. I suspect there must be other people quietly experimenting with it.

I’d be particularly interested in speaking to anyone modelling horse racing, betting exchanges or other sports with similar structures. I’d love to hear from people working on variable-sized ranking problems, market prediction or large chronological tabular datasets etc

I’m not expecting anyone to reveal proprietary data 😉