Starting points for implementing ML/AI into sports trading

[Gcampb]

I have developed several machine-learning models for horse-racing markets. They have now been live, rather than in testing, for around seven months, so I am reasonably confident that their real-world performance is broadly in line with what I observed during the training and validation period, which used data from 2015 onwards.

I also try to work closely with others involved in similar projects, and I provide selections for a parimutuel syndicate, so I feel able to speak with some confidence on the subject.

One point that is often misunderstood is that using a generative AI system is not the same thing as building and validating a machine-learning model for statistical prediction. They are fundamentally different approaches.

A major problem with using an AI model for systematic statistical analysis is validation. Unless the entire AI-based process has been recreated and tested across several years of historical data, it is very difficult to know whether the results are genuinely reliable or whether a short run of apparent success is simply noise.

I have yet to see a horse-racing AI project that has been properly backtested over a large historical sample. One likely reason is the cost and complexity involved. Whatever data preparation, prompts, context and query process are being used live would also need to be replicated historically. That could mean tens of thousands of model queries and hundreds of millions of tokens, potentially at significant financial cost.

Any meaningful change to the prompts, data or methodology could then require the entire test to be run again. I have not heard of anyone going to that level of effort with a generative-AI racing system.

There is also a more basic question. If you are prompt-engineering an AI model to analyse a fixed set of data points and produce a consistent, systematic output, why use a generative AI model at all? In many cases, the same logic could be implemented programmatically, with lower cost, greater consistency and much easier testing.

With conventional machine learning, the process is different. Historical testing, calibration and validation are integral parts of model development. A properly designed validation process can give you a reasonable expectation of how the model should perform on genuinely unseen data. You can then run it in shadow mode on live races and compare the observed results with the expected performance before committing capital.

That does not mean machine learning automatically produces an edge. Poor data, leakage, overfitting and weak validation can make an ML model look far better than it really is. However, the framework at least allows those risks to be measured and tested systematically.

Even then, the predictive model is only half the battle. Finding selections with an apparent statistical edge is one problem; executing those selections efficiently in a live market is another. Entry timing, liquidity, slippage, stake size, price movement and trade management can determine whether a theoretical edge survives in practice.

In my view, execution is at least 50% of the challenge.

Edit: I just reread OP’s original question and realised I haven’t quite answered it - excuse the rant above - but I wanted to be clear about the ml vs ai approaches.

If you actually want to start working on a ML project, the obvious starting point is getting a dataset and more importantly a data pipeline. In this last year there have been several horse racing datasets pop up in various places - which will save you a lot of time having to scrape the historic results - this isn’t a bad starting point :https://www.kaggle.com/datasets/deltaro ... -2025/code

But, in order to have a live working model after training you’ll need to establish a way to keep the base dataset up to date - which means you need to get comfortable with web scraping.

That dataset is a starting point - there are other bits and pieces you will want to put together, weather data perhaps, sectional data (this is actually only reliably available from 2019 onwards and is patchy, so it’s not something you need in your first models). I use multiple sources of data, but that kind of constitutes what the secret sauce is. Get creative!

The next thing you’ll need to think about is Features. You have your raw data - but that’s pretty useless on its own. Feature engineering is the single biggest point which will make or break the project. This is where domain knowledge is key. Personally, I have 9 seperate models I use for scoring a race card and some of those models use up to 2500 individual features. More is not necessarily better however through ablation testing you can remove and refine features for models. If you have a good understanding of horse racing the chances are you are already ahead of 99% of the PHD quants who don’t have that base knowledge of the sport to know where to begin with feature engineering in horse racing.

Training and testing - this is most tedious part, you need to be really careful here. It’s also where you decide what question you want answered - do you want to get win and place probabilities, that’s a good start but there are some more interesting things you can try to answer depending on what you plan to use the model for. It would be worth looking into different types of machine-learning models, such as logistic regression, decision trees, random forests, CatBoost, XGBoost and neural networks. They each learn patterns differently and have different strengths depending on the type and size of the data. But you need to decide how you will train them, how you will tests them and always leave room for final unseen tests. It’s all about experimentation and benchmarking to find what works best with your features. One warning I’ll give - if the results are too good to be true - you have leakage of future data leaked into the features.