How Accurate Are AI Horse Racing Picks?

20 June 2026 | Updated 22 July 2026 | Realtime Tips

Most people asking this question want a simple answer: how often does the AI pick the winner?

That is why, for this article, we are using win strike rate as the main measure of accuracy. It is not the only useful measure in betting - profit, ROI, price discipline and whether a selection beats market expectation all matter - but win strike rate is the clearest way to answer the question most searchers are really asking.

In our current dataset, GPT-5.4 Advanced Tips selected 1,097 winners from 3,752 races, giving an all-selections win strike rate of 29.2%.

For context, the current Racing Post Press Challenge table shows The Guardian on 1,090 winners from 4,013 tips, listed at 27%. On an exact-count basis, that is 27.16%. Comparing all selections with all selections, GPT-5.4 Advanced Tips was therefore 2.08 percentage points higher, or about 7.6% higher on a relative win strike-rate basis.

That is the fairest headline comparison. We also found stronger historical results after applying confidence-band filtering, but those need to be treated separately.

The short answer

AI horse racing picks can be accurate when the model is not treated like a generic chatbot.

In our testing, the best results came when the model was given a structured race dossier and instructed to analyse the supplied evidence: form, going, course, distance, draw, trainer and jockey data, ratings, prices, previous runs and race context.

Our current GPT-5.4 Advanced Tips results are:

What does accurate mean for AI horse racing tips?

There are several ways to judge a racing prediction:

• Win strike rate: how often the selected horse wins.
• Place rate: how often it finishes in the places.
• Profit and loss: whether the selections return more than they stake.
• ROI: profit as a percentage of stakes.
• Value: whether the selection was better than the market implied.
• Beating SP or BSP: whether the model identifies prices that shorten or outperform expectation.

For a betting system, win strike rate alone is not enough. A 20% strike rate can be profitable at the right prices, while a 40% strike rate can still lose money if the prices are too short.

But for the question "How accurate are AI horse racing picks?", win strike rate is the most natural starting point. It answers the plain-English question: how often does the AI pick the winner?

Our current Advanced Tips result: 29.2% from 3,752 races

Across all GPT-5.4 Advanced Tips selections in this dataset:

• Races analysed: 3,752
• Winners selected: 1,097
• Win strike rate: 29.24%

That means roughly three winners from every ten races, across all selections, before applying confidence-band filtering.

This matters because it avoids cherry-picking. If we want to compare Advanced Tips with human newspaper tipsters, the fairest comparison is not our best filtered band against all of their picks. It is our all-selections record against their all-selections record.

How does that compare with newspaper tipsters?

The Racing Post Press Challenge is a useful public benchmark because it tracks newspaper and tipster selections across the year. The table records wins, tips, percentage of favourites tipped and £1 stakes at SP. Racing Post explains that competitors are measured from a theoretical £1,000 bank, with £1 staked at SP on qualifying races.

At the time of writing, the leading newspaper-tipster benchmark by win strike rate in that table is The Guardian. Because the public table updates, these figures should be treated as a snapshot checked on 20 June 2026:

• The Guardian: 1,090 winners from 4,013 tips
• Exact strike rate: 27.16%
• Published table percentage: 27%

By comparison:

• GPT-5.4 Advanced Tips: 1,097 winners from 3,752 races
• Exact strike rate: 29.24%

So the like-for-like comparison is:

GPT-5.4 Advanced Tips: 29.24%
The Guardian benchmark: 27.16%
Difference: +2.08 percentage points
Relative uplift: +7.6%

A careful way to say this is:

A less careful way would be to say "AI is 7.6% better than newspaper tipsters." We would avoid that wording. The evidence supports a more specific claim: our GPT-5.4 all-selections win strike rate was around 7.6% higher than that benchmark snapshot.

What about the 41.2% confidence-band result?

We also found a much higher strike rate when applying confidence-band filtering.

When we ran scenarios across roughly 3,700 GPT-5.4 selections and selected the most profitable confidence bands, the trimmed high-confidence subset produced:

• Selections in selected confidence bands: 675
• Win strike rate: 41.2%

That is a strong historical result, but it should not be used as the main comparison with newspaper tipsters.

Why? Because the newspaper-tipster benchmark is based on all selections, whereas the 41.2% figure comes from a filtered subset chosen after analysing the historical performance of confidence bands.

The right interpretation is:

That distinction is important. It is the difference between honest analysis and overclaiming.

How Advanced Tips actually works

Advanced Tips does not simply ask an LLM, "Who wins this race?"

Instead, each race is converted into a structured race pack from our database. The model receives two main pieces of instruction:

1. System prompt: the model's standing job description. It tells the model to act as an expert horse racing analyst, use the supplied evidence, consider form, going, course, trainer, jockey and value, and return output in a format we can parse.
2. User prompt: the race dossier. For each race, we build a fresh prompt containing the race details, every runner, runner IDs, stats, previous runs, trainer and jockey data, draw, going, surface context, prices and the exact analysis requirements.

The key point is that Advanced Tips is designed to make the LLM perform structured analysis over supplied racing data, not invent a racing opinion from memory.

That matters because a language model on its own can sound confident even when it is wrong. A language model constrained by race data, runner IDs, structured fields and post-response validation becomes much more testable.

Case study: Mission Central at Royal Ascot

A good example came in the 15:40 Royal Ascot, Tuesday 16 June 2026, in the King Charles III Stakes (Group 1).

The race conditions in the prompt were:

• Trip: 5f
• Surface: turf
• Going: good to firm
• Field size: 26 runners
• Draw: no significant draw advantage identified

The GPT-5.4 Advanced Tips selection was:

• Horse: Mission Central
• Runner ID: 4205777
• Profile: 3yo gelding
• Trainer/jockey: A P O'Brien / Ryan Moore
• Forecast price in prompt/export: 15.0 decimal
• SP: 15.0 decimal, or 14/1 fractional
• Result: won

This was not a favourite-following pick. Mission Central was around seventh in the market. The more obvious market or basic-stat selections were elsewhere: Overpass was favourite, Night Raider was prominent in the betting, Big Mojo had the top official rating and Asfoora was another highly rated contender.

GPT-5.4's reasoning was deliberately measured:

The model gave it a small stake, at 1.0%, with 12% confidence. In the workbook summary, that 1% win stake returned 0.15 units, for +0.14 units profit.

What made the pick interesting was not just that it won. It was how the model got there. It did not simply sort by odds or official rating. It elevated a horse with:

• rapid recent progression
• back-to-back 5f wins at Naas
• a last-time-out win
• a 71.43% win strike rate
• a 50% distance win strike rate
• a 3yo weight allowance against older rivals
• previous Ascot-winning form on good ground
• the Aidan O'Brien and Ryan Moore combination

It also respected the difficulty of the race. A 26-runner Group 1 sprint is not the place for false certainty, so the model picked the winner while still recommending a small stake and modest confidence.

That is the type of behaviour we want from AI racing analysis: not just a selection, but a traceable reason for the selection and a realistic view of the uncertainty.

Why structured data matters

The difference between a useful AI racing pick and a generic AI racing opinion is the quality of the evidence.

A weak AI prompt might ask:

That encourages the model to guess, generalise or rely on incomplete context.

A stronger prompt gives the model a structured dossier and asks it to reason from the supplied data. That means the model can compare each runner on relevant racing factors:

• recent form
• suitability of going
• course and distance evidence
• trainer and jockey records
• draw and pace context
• official ratings and other performance indicators
• market prices
• historical comments
• field strength
• value and confidence

In other words, the model is not being used as a magic tipster. It is being used as an analysis layer over structured racing data.

The main risks: hallucination and false precision

AI racing predictions have two big failure modes.

The first is hallucination risk. A model can produce a confident-sounding reason that is not supported by the data. In horse racing, that is especially dangerous because the language of racing analysis can sound persuasive even when it is wrong.

The second is false precision. A confidence score can look scientific even when it is only an estimate. Saying a horse has a 12% chance does not mean the model knows the future to two decimal places. It means the model has assigned a probability based on the evidence and the prompt design.

That is why Advanced Tips is designed with safeguards.

Safeguards we use in Advanced Tips

1. Supplied-data grounding

The model is told to base its conclusions on the race data we provide: runners, IDs, form, going, course, distance, draw, trainer and jockey records, ratings, prices and historical comments.

This reduces the chance of the model inventing outside facts.

2. Mandatory runner IDs

Every selection must include the exact runner_id from our database.

That prevents vague tips like "the O'Brien horse" and lets us validate the response against the actual runner record.

3. Structured output

The model must return parseable fields such as:

• Predicted winner
• Bet advice
• Bet percentage
• Confidence
• Full runner ranking

If the output shape is wrong, it does not flow cleanly into the audit and export pipeline.

4. Confidence calibration

We tell the model to treat confidence as a real win probability, not marketing language.

In the current concise prompt, every runner gets a probability and the probabilities should add up to roughly 100%. That helps prevent the model from giving unrealistically high confidence to several horses in the same race.

5. Low-confidence signalling

The model is not forced to turn every race into a strong bet.

It can return:

• Avoid
• Small stake
• Each-way
• Win

Mission Central is a good example. GPT-5.4 picked the winner but still marked it as a small stake at only 12% confidence because it was a deep, competitive Group 1 sprint.

6. Database validation after response

The parsed runner ID is joined back to our runner and result database, including:

• finish position
• SP and BSP where available
• forecast price
• rating
• trainer
• jockey
• non-runner checks

That is how we audit whether the pick was a real runner and how it performed.

7. Multiple-model comparison

The daily pipeline can run separate high-effort outputs from models such as GPT-5.4, GPT-5.5 and Claude variants.

That lets us compare agreement and disagreement. The ensemble process can flag consensus, divergent opinions and confidence differences rather than hiding them.

8. Stake and price discipline

Downstream workbooks apply bet advice, confidence bands, price filters and Kelly-style staking logic.

So even when the model likes a horse, the staking layer can scale the bet down or exclude it.

None of this makes hallucination impossible. The point is to make every prediction traceable, parseable, validated and auditable.

That is the difference between "AI wrote a tip" and "AI produced a testable prediction."

What the results do and do not prove

The current results are promising, but they should be interpreted carefully.

They do show that:

• GPT-5.4 Advanced Tips achieved a 29.2% all-selections win strike rate across 3,752 races.
• That all-selections strike rate is higher than the leading newspaper-tipster benchmark we compared against.
• Confidence-band filtering found a historical subset with a 41.2% win strike rate.
• A structured LLM approach can identify non-obvious winners, such as Mission Central at 14/1.

They do not yet prove that:

• the 41.2% banded strike rate will hold in live forward testing;
• every future model version will perform the same way;
• AI tips are automatically profitable;
• win strike rate alone is the best measure of betting value;
• AI picks should be followed blindly.

The next stage is forward testing: pick the bands, lock the rules, then measure future races that were not used to select those bands. You can follow that work on the Advanced proof page.

So, how accurate are AI horse racing picks?

Based on our current GPT-5.4 Advanced Tips dataset, the answer is:

The stronger filtered result is:

The broader lesson is simple: AI horse racing picks are only as useful as the data, structure and testing behind them.

A generic chatbot opinion is not enough. But an LLM guided by structured race data, validated runner IDs, probability discipline and outcome auditing can produce predictions that are both useful and measurable.

That is what we are building with Advanced Tips.

Sources and notes

• Realtime Tips internal Advanced Tips dataset: GPT-5.4 selections, 3,752 races, 1,097 winners, 29.2% strike rate.
• Realtime Tips internal confidence-band analysis: 675 selected GPT-5.4 banded selections, 41.2% historical win strike rate.
• Racing Post Press Challenge snapshot checked 20 June 2026: The Guardian listed at 1,090 winners from 4,013 tips, shown as 27%; The Favourite listed separately at 37% with 100% favourites tipped. Source: Racing Post Press Challenge.
• Mission Central case study verified against public race reporting and result pages, including Racing Post coverage and the Sporting Life result page for the King Charles III Stakes on 16 June 2026.

Racing predictions are uncertain. Only bet what you can afford to lose and read our responsible gambling guidance.