AI Horse Racing Predictions: Can AI Pick Winners? An Ex-Quant's Verdict

Home › Betting Strategy › AI Predictions

By The Guv'nor · Last updated October 2026

Can AI predict horse racing? Partly. A well-built model can rank the runners and put a fair chance on each, but the betting market already does that very well. The real test of any AI horse racing predictions is whether they know something the prices don't, and most AI tipsters and chatbots rarely show evidence that theirs do.

Can AI predict horse racing? The short answer

Look, this is my home ground. I was a Quantitative Researcher at a Tier 1 investment bank in the UK, essentially a statistics/maths geek. So when people ask me about AI horse racing predictions, I don't ask "is it clever?". I ask "is it better than the odds?".

A computer can predict horse racing in the sense that its top-rated horses win more often than its outsiders. So can the bookies' prices. The only bit that pays is spotting the races where the model and the market disagree and the model is the one that's right.

How horse racing prediction algorithms actually work

Almost every racing model does the same job: it takes the form book and gives each horse a win probability, adding up to 100% in each race.

The classic method, written up by Ruth Bolton and Randall Chapman in 1986 and later used by William Benter in Hong Kong, is a multinomial logit model. It learns from past races how much each factor matters: Benter's list includes finishing positions, lengths beaten, normalised times, weight carried, strength of opposition, the draw, jockey ability and distance preference.

Modern "AI" products may use fancier machine learning, but the inputs are the same public form you can read yourself (see how to read horse racing form for what those figures mean).

From a probability to a bet

A probability isn't a bet until you add a price. A worked example, not a real race:

• The model gives a horse a 20% chance. Fair odds for 20% are 4/1 (5.0 in decimal).
• At 5/1 (6.0), the expected return is 0.20 × 6.0 = 1.20: +20% per pound staked, if the model is right.
• At 3/1 (4.0), it's 0.20 × 4.0 = 0.80: a 20% loss on average.

Same horse, same model, opposite answers. That's why value betting in horse racing matters more than any prediction.

Where ChatGPT fits in (it mostly doesn't)

Chatbots are language models, trained on text to write plausible answers, not fitted to race results to produce probabilities. Ask for ChatGPT horse racing tips and you get a confident paragraph about "strong recent form" that may be stale or wrong. Even with web search on, it's summarising other people's views, not pricing the race.

What the best-documented test of a racing model found

The best-documented account of a real-money racing model I know of is William Benter's 1994 paper "Computer Based Horse Race Handicapping and Wagering Systems: A Report", in the book Efficiency of Racetrack Betting Markets. He ran a computer betting operation in Hong Kong, and two of his tables, rebuilt below, tell you more about AI tipsters than any sales page.

Method note: all figures are copied from the paper. The data is every race run by the Royal Hong Kong Jockey Club from September 1986 to June 1993: 3,198 races and 32,877 runners. The market here is the Hong Kong tote, not UK bookmaker or Betfair prices, so treat this as evidence about how models behave, not a UK edge to copy.

Table 1: the model was overconfident on exactly the horses it wanted to back

Benter's model against what actually happened, only for runners the model rated higher than the market did: the horses a model-follower would have bet.

Source: Benter (1994), Table 4. Every single band won less often than the model said, and in the busy 5-15% bands the shortfall is more than four standard errors, so it isn't bad luck. Benter called the bias "extreme and consistent", with the actual results always pulled in the direction of the public's estimate.

So the horses where a model disagrees with the market are exactly where it's overconfident. A model that looks well calibrated across all runners can still lose you money on the ones it tells you to back.

Benter's fix was a second step: blend the model with the market's probabilities. The bias then disappears. For the same horses, runners the combined model put at 5-10% averaged 7.2% and won 6.9% of the time (4,511 runners), and the 15-20% band said 17.3% and won 17.4% (1,426 runners). Source: Benter (1994), Table 7.

Table 2: a tipster consensus added almost nothing to the market

Benter scored each predictor with a pseudo-R², where 0 is no better than random guessing and 1 is perfect. What matters is the last column: how much a predictor improves on the market's odds when you combine the two.

Source: Benter (1994). Samples: September 1986 to June 1993 (3,198 races) and September 1988 to June 1993 (2,313 races).

On their own, the 9-factor model and the tipster consensus scored almost the same. Combined with the market, the model added something and the tipsters added next to nothing. Benter's conclusion: when the tipsters and the public disagreed, the public's estimate was superior.

And notice how small the scores are: 0.14 at best on a scale running to 1. Racing is noisy. A good model is right about probabilities, not winners, which is why nobody can promise you a winner every day (anyone that can is a fraudster).

How to test an AI tipster or a ChatGPT prompt yourself

If someone's selling AI horse racing tips, or you fancy trying a ChatGPT prompt, test it properly first:

1. Record before the off. Log every pick with the time and the price available. Results quoted afterwards are worthless.
2. Use prices you could actually get. If the record uses Betfair SP, label it BSP and remember commission comes off.
3. Level stakes, every pick. One point per selection, no picking and choosing afterwards.
4. Don't test a chatbot on old races. It may already have read the result during training, so a "past race" test flatters it. Only races run after you ask count.
5. Wait for a real sample. A few dozen bets proves nothing, especially at bigger prices.
6. Compare with the market. If backing the favourite did as well, the AI adds nothing.

I won't print a table of AI picks here until I've run that test properly on races run after the picks were made. A made-up "I tested ChatGPT" table is exactly what I'd warn you about. For judging any paid service, see my guide on whether horse racing tipsters are worth it.

The market is the model to beat

Here's the bit the AI sales pages skip: the odds are a prediction too, made by everyone with money in the market. Benter put it well: trainers' and jockeys' intentions, secret workouts and "whether the horse ate its breakfast" are known to certain parties, and their betting shows up in the odds. A model built on published form always misses that.

So a racing model competes with the price, and the price has a head start: in Table 1, on the very horses the model liked, the results were always pulled back towards the market's view. That's why the professional approach blends the two.

For how UK and Irish prices line up with results: horse racing odds explained compares actual win rates with the chances implied by the odds, and how often favourites win breaks it down by race type and field size.

Overfitting and small samples: why most AI systems fall apart

Overfitting is when a model learns the noise in its training races instead of the signal. Give a computer enough factors and it will always find a pattern that "would have" made money. Benter warned that "some overfitting will always occur" and tested on races the model never saw.

To show how easy it is to fool yourself, I ran a simulation (a worked example, not real betting data).

Simulation: 200 "systems" with no edge at all

Every bet is at 5/1 on a horse with a true 15% chance, so every system loses 10% long term. Each system gets 150 bets; I pick the best, give it 150 fresh bets, and repeat the whole thing 2,000 times.

Nearly a quarter of zero-skill systems look like winners after 150 bets; the best shows a 40% profit. Then it goes straight back to losing 10%, because it never had an edge. Its 24.9% chance of a second profitable run is no better than any random system's (24.2%).

An AI tipster showing you the best of thousands of backtests is showing you that top row. The horse racing betting systems guide covers the small-sample trap.

Benter also noted that overestimating your edge by a factor of two is "easily done in practice", and that overestimating it by more than that makes Kelly staking shrink your bank. The staking side is in my guide to staking plans for horse racing.

How I use numbers in my own process, and what a human still adds

If I am honest, I'm not anti-model. My stats skills are what made my betting profitable enough that most of your favourite bookies won't take a bet from me.

The outsiders in my newsletter come from a new but proven system that picks out horses that have won or come close at a higher weight or mark than they're running off today. That's a model in all but name. But the system doesn't send the newsletter: I filter through what it finds so you don't have to go through every racecard, and I bet on every horse I tip.

On top of that comes the stuff Benter said a model can't see. I own racehorses, so now and then I pass on stable information, plus "smart money" moves in the market as they happen. My 'Banker' bets are simply the shorter-priced, solid ones; with the outsiders, even when they only place we make more profit than an odds-on shot winning.

The discipline matters more than the algorithm. Strict points staking, say 1pt = £5 on a £100 bank (see what 1pt means in betting). No chasing, no going all in on a feeling, and skip any bet you're unsure about. Some days there's nothing worth sending: a stinky day for the ponies.

Losers happen, model or no model. I judge myself on profit month to month, not win rate, and that's how you should judge any AI tipster too.

My bets come from my own numbers plus the form study a chatbot can't do; members get them in the morning newsletter, a short list for your betting of the day. See how my premium tips work, or read the rest of my horse racing betting strategy guide.

AI horse racing predictions: your questions

Can ChatGPT predict horse racing?

Not in any way you should bet on. ChatGPT is a language model, not a probability model fitted to race results, and it doesn't price a race against the market. Its "picks" can rest on stale or wrong information. If you try a ChatGPT horse racing prediction, test it on races run after you ask, at the prices on offer.

Is there a free AI horse racing predictor for UK racing?

Plenty of apps and sites call themselves an AI horse racing predictor for the UK. I won't name or rate any because I haven't tested them under the rules above. Ask for a time-stamped record at prices you could have got, over hundreds of bets, before you trust one.

What is the best horse racing prediction algorithm?

The well-documented starting point is the multinomial logit model described by Bolton and Chapman in 1986, which Benter built on. Fancier machine learning can be swapped in, but the evidence above says the algorithm matters less than two things: testing on races the model hasn't seen, and blending its output with the market's odds.

Can AI predict horse racing results accurately?

Not in the sense of naming winners most of the time. Even Benter's combined model scored 0.14 on a 0-to-1 scale. A good model is accurate about chances: its 20% shots win about one race in five over a big sample. Whether that makes money depends on the price.

More in this guide

• Horse racing betting strategy: 10 rules from an ex-quant
• Value betting in horse racing
• Horse racing betting systems that work? I tested the popular ones
• Are horse racing tipsters worth it?
• How often do favourites win?