From James · Opinion
The bookies aren’t afraid of AI.
They should be rubbing their hands together.
AI is becoming astonishingly good at explaining why a horse might win. That may be exactly why bookmakers have very little to fear.
A friend asked me a question the other day, in the pub, with the racing on in the corner and neither of us paying it much attention. With AI getting this powerful, are the bookmakers worried about what happens when it can predict horse racing?
It is a completely reasonable question. A machine that can read a million data points before you have finished your sentence, pointed at a sport that produces data continuously and has done for two hundred years. You can see why people assume the layers are nervous.
My answer surprised him. I think the bookmakers will be rubbing their hands together.
Not because artificial intelligence is useless. The opposite. Because I suspect it is about to make a very old betting problem more persuasive than it has ever been.
The bookies may love this
Here is what a betting business actually sells. Not outcomes. Confidence.
A bookmaker does not need you to be wrong. A bookmaker needs you to bet, and then to bet again next Saturday. The most valuable customer in that model is not the one who loses. It is the one who is certain.
For most of racing history, certainty was expensive. You built it yourself, slowly, out of form books and late nights and a decade of being wrong in instructive ways. Most people never got there, which is precisely why most people bet small and occasionally.
Now anybody can have four paragraphs of fluent, structured, entirely reasonable-sounding analysis of the 3:40 at Wolverhampton in about six seconds. It will mention the going. It will mention the trainer's strike rate at the track. It will sound like somebody who knows.
I think that is worth more to the betting industry than any pricing model they could build themselves.
AI has inherited an old betting problem
I want to be careful here, because the lazy version of this argument is that AI just parrots what humans already said, and that is not true and not interesting.
These systems can genuinely analyse information at a scale no person can approach. That is real. The trouble is not the analysis. It is the inventory.
Almost all the racing information available to a consumer AI is information the wider market can also see. The same form, the same ratings, the same going reports, the same trainer statistics. Which means tens of thousands of people, and rather more importantly a market with actual money in it, are looking at the same material and have already moved on it.
There is information, and then there is information the market has not priced. Only the second one pays.
A price is not somebody's opinion. It is everybody's opinion, weighted by how much they were willing to stake on it. To beat that you do not need to be right. You need to be right about something it has missed.
We went looking for the prediction engine too
I should be honest about this, because it would be very easy to write the above as though I had always known it.
Elite Pass started as an attempt to find the horse. Better ratings, cleverer weighting, sharper reading of the shape of a race. Over the life of the programme I have examined 227 systems, rules and racing theories. Zero of them have cleared our validation bar. Every one of those failures is still published on the site, named, with the test that killed it.
The most useful failure was also the simplest. We looked at 60,671 races with one clear favourite. The favourite won 35% of them and lost 65%. Which is, to a close approximation, exactly what its price had already said it would do.
The market was not wrong. We had been asking the wrong question.
Being right isn't enough
Suppose your AI is genuinely excellent. It reads the race properly, weighs everything correctly, and tells you the favourite should win. And the favourite wins.
Did you make money?
Only if the price was wrong. If the market already had that horse at the same probability your model landed on, you have been paid precisely what the risk was worth, minus the margin. Do that a thousand times and you will finish slightly behind, having been right almost every time.
The prediction can be excellent. The bet can still be terrible.
That distinction never fits comfortably in a headline, which is part of why it keeps getting lost. Accuracy is intuitive. Value is not.
The most dangerous prediction is the convincing one
A bad tip is survivable. You read it, something in you objects, you leave it alone.
A well-argued one is a different animal.
Give any model enough variables and it will find a subset that looks extraordinarily predictive. Course, distance, going, class, month, field size, market position. Slice finely enough and something will glow, and a good deal of that glow is the searching itself, not the sport.
Then the machine writes it up, and it writes it up beautifully. Clear structure. Calm, confident tone. Citing the very features that made the subset look special in the first place. It reads like insight because it has all the surface properties of insight.
And the horse, unfortunately, hasn’t read the analysis.
So is AI useless for racing?
No. I think we are pointing it in the wrong direction.
Ask a model who wins and you have put it in the business of manufacturing an opinion. It will produce one, because that is what you asked for, and it will not tell you that the honest answer was a shrug.
AI is potentially much more interesting as a sceptic than as a tipster.
Ask it instead what would have to be true for a price to make sense, what is missing from the picture, what genuinely cannot be known before the stalls open, and it becomes something quite different. A fast, tireless, completely unembarrassed devil's advocate. It will never get bored of being asked to find the hole in an argument. No human analyst can say that, least of all about their own work.
There is another experiment worth trying
The single most useful instruction I have found is this one:
Now make the strongest possible case for why your own prediction is wrong.
What comes back is usually more valuable than the prediction was. The model will surface the assumption it quietly leaned on, the run it treated as representative when it was not, the bit of the race it cannot see. Occasionally it will dismantle its own answer so effectively that you put your wallet away, which I would argue is the single most profitable thing a piece of racing software has ever done for anybody.
That is roughly where our own thinking ended up. We stopped asking which horse would win and started asking three other things instead. What does not fit? Where is this race fragile? What does the market already know?
Sometimes the honest output is that we cannot read the race at all. We publish that too, and call it No Read, and it is a genuine answer rather than a failure to produce one. We are becoming increasingly interested in teaching machines to doubt.
The bookmakers don't need AI to be wrong
Which brings me back to my friend in the pub.
Everyone frames this as a race between the models and the layers, and assumes that if the models get good enough the layers are in trouble. I do not think that is the shape of it at all.
The bookmakers don’t need AI to be wrong. They need it to be persuasive.
A generation of bettors carrying an articulate, endlessly patient, always-available reason to have a bet strikes me as extremely good news for the people taking the other side of it. Not because the analysis is bad. Because confidence and edge are different things, and only one of them is being mass produced right now.
I should say plainly that this last part is opinion. It is a prediction about human behaviour, not something I have measured, and I would be delighted to be wrong about it. Everything above it about prices and information is a good deal more solid, and most of it is on the site with the workings attached.
Try this yourself
Take any race today, open whichever assistant you use, and run these three in order. It takes about two minutes.
- 1
What are the strongest evidence-based reasons the favourite could underperform today?
- 2
Which assumptions would need to be true for the favourite’s current price to be justified?
- 3
Now argue against your own conclusion, and identify anything in the data that cannot reliably be known before the race.
Don’t treat the answer as a tip. Treat it as an experiment. You are not looking for a winner. You are looking at how quickly a confident argument comes apart when you ask it to.
The third prompt is the one that does the work. Most people never run it, because by then they already have the answer they wanted.
We stopped asking which horse would win.
We started asking what makes a race unstable.
James Kennedy
Founder, Elite Pass
Related research
- What two months of testing actually taught usThree of the first five beliefs we tested did not survive. Here is what lived, what died, and what changed about the question we were asking.
- We looked at 57 races and 41 got no readOn one race day the read assessed 57 races, scored 16 and refused 41. Exactly one race went the whole way. Here is what refusing a race actually involves.
- Big fields and what they do to a favouriteMore runners means more ways to be beaten. That part is true. The awkward part is that the market worked it out before we did.
Elite Pass publishes research, not guaranteed outcomes. Findings are measured against the market and remain subject to replication and sample size. The evidence room and the public ledger carry the full record.