Are the AI's most-confident picks profitable?

This is the most seductive entry on the whole board, and you can see why in one sentence: back the machine only when the machine is sure.

The pitch sells itself. We built an AI model that prices up every British race, runner by runner, off form, conditions and weather. Most of the time it is hedging, spreading its opinion across a field where it has no strong read. But every so often it plants a flag. It rates one horse 35 per cent or better to win and effectively says, this one I am confident about. The claim is that those flagged runners, the strong fancies, are where the money lives. Ignore the model's noisy guesses, wait for the big confident picks, pile in, and let the clever computer do the hard work.

It is the cleanest logic on the page. A confident filter feels like it should strip out the rubbish and leave only the gold. The model is well calibrated, and the headline number, when you finally see it, lands a whisker from break-even. To a hopeful eye that reads as basically free money waiting for one small tweak.

So we tested it honestly, the same way we test every system in the Lab. Flat stakes, real prices, every confident pick settled as it actually finished. The question is simple and it is the one every punter who has ever trusted a tipster wants answered. When the smart model is genuinely sure, can you make money following it? The result is the closest any system came to passing. It still did not pass.

The appeal runs deeper than the other systems we test, because this one is not obviously daft. Doubling your stake after a loss is a maths trap you can explain in a minute. Backing only an AI's strongest fancies sounds like exactly what a sensible person should do.

The selectivity flatters it. When you throw away the model's worst longshot guesses and keep only the confident shortlist, you are left with well-fancied, short-priced runners that win a lot. Watch them go in race after race and it feels like proof. The strike rate is high, the losers are rare, and the running total drifts along close to flat instead of sliding away the way a random or longshot strategy does. Nothing about the experience screams losing system, which is precisely the problem.

Then there is the number itself. A return of -0.8% reads, to a hopeful eye, as nearly winning. People look at a figure that close to zero and assume the gap is a rounding error or a small fee, the sort of thing one clever adjustment would erase. They are already spending the profit that is not there.

Underneath it all sits a single confusion, and it is the most important idea on the page. People run together two completely different things. One is a calibrated model knowing which races it has a genuine read on. The other is a model that can turn that read into profit. The first is real and the AI genuinely has it. The second is not. Pascal looks at -0.8% and hears it as nearly winning, a machine that is almost there. The Professor looks at the same number and sees a system that is simply losing slowly.

Here is the genuinely interesting reason the confident picks land near break-even, and it is not the reason people hope.

When the model is genuinely sure about a horse, it is almost always agreeing with a short price the market has already got roughly right. That matters, because of where the bookmaker's margin sits. Short prices carry the thinnest slice of the overround. Odds-on favourites come back at -4.6%, against -34.5% for outright outsiders. A basket of confident, shorter-priced runners inherits the least-padded corner of the market. That is the whole secret. It is not skill. It is renting the least expensive seat in the house.

The catch is fatal to the dream. The model has no market price in its features, no speed figures and no Betfair starting price. It cannot out-read the crowd on these races because it is reading the same form the crowd already read. By the time the machine and the market both agree a horse should win, the bookmaker has already shortened it to match. There is no spare value left to take. The model is echoing a price, not beating it.

So the arithmetic closes in. The overround is baked into every bet, about 12% per race, climbing towards 30% in big fields. Even the sharpest, shortest runners give a little back on average. A steady -0.8% across hundreds of bets is still a loss, just a polite one. The confident filter buys you the cheapest corner of a priced-up market, and the cheapest corner of a losing bet is still a losing bet.

We ran this the same brutal way we run everything in the Lab, with no thumb on the scale for the AI.

The pool is 27,421 real British races. For the confident system we kept only the runners the model rated 35 per cent or better to win, the picks where the machine plants its flag. That left 361 bets. Every one was staked flat, the same notional amount, so a lucky big-priced winner cannot quietly carry the whole record. Then every bet was settled to industry Starting Price, the odds you would actually have got walking up at the off, not some flattering forecast price.

The settling rules are where most published backtests cheat, so we are blunt about ours. Fallers count as losing bets. Pulled-up horses count as losing bets. A horse that was sent off your confident pick and unseated at the second is a loss, full stop, because that is what it costs you in real life. Joint-favourites are split rather than double-counted. No Betfair commission is charged, which if anything makes the system look better than reality, not worse.

That last point matters here more than on any other page. This is the one system that brushes break-even, so the way you settle it decides everything. Settled honestly, to industry Starting Price, with non-runners voided, fallers and pulled-up horses counted as the losing bets they are, every race required to have a recorded winner and joint-favourites split, the confident picks come in at -0.8%. Settle them any softer than that and you can flatter the number into a fiction of profit, which is exactly what every tipster on the internet is selling you. We would rather publish the honest minus sign.

Here is the result, plainly. Across the backtest the AI's most-confident picks returned -0.8% to Starting Price over 361 bets. Read that honestly. It is a loss. It is a very small one, and it sits inside the error bars of break-even rather than proving any edge, but it is still a minus sign, not a profit.

The strike rate is high, and that is the part that fools people. These are short-priced, well-fancied runners, so a healthy share of them win. But a high strike rate on short prices is exactly what break-even looks like. You win often, you collect small returns each time, and the overround skims its cut off every price, so the total still drifts gently into the red. Winning most of your bets and still losing money is not a paradox. On short prices it is the default.

Now the honest health warning on the -0.8% itself. The figure is fragile and easy to flatter. The confident sample is small at 361 bets and heavily skewed towards National Hunt, so any brush with zero is happening on a thin, jumps-heavy slice of history. The 95 per cent range around the number comfortably includes both a small profit and a small loss, which is another way of saying the result is statistically indistinguishable from zero. Knock the rounding either way and you are still hovering at break-even, with the honest reading a slight loss.

And this is to Starting Price with no commission charged. Add real Betfair commission, or simply run the same confident filter forward on a normal mix of fresh races, and the wafer-thin cushion is gone. Out of sample it reverts to a loss. There is no number here you can build a stake on.

So, are the AI's most-confident picks profitable? No. They hold up better than any system we have measured and are still not a way to make money.

A return of -0.8% to Starting Price over 361 real races is a small, fragile loss that lives inside the noise of break-even, and it only ever brushed zero on a thin, jumps-heavy past sample to SP with no commission charged. It suits nobody as a staking plan. Charge real Betfair commission, or just run it forward on a normal mix of races, and the cushion vanishes. Treat the -0.8% as a backward-looking, no-commission curiosity, never as a wage and never as a reason to bet.

But do not throw the model out with the bathwater, because the real lesson here is the most useful thing on the whole board. The confident filter proves two true things at once. It proves the AI knows which races it can read, because by refusing to bet its worst longshot guesses it dodges the -11.4% loss of backing its own top pick blind, a hole deeper even than the -8.7% you take blindly backing the favourite. That is genuine insight. And it proves the same AI cannot beat an efficient market, because it has no price, no speed figures and no Betfair SP to beat it with. A model knowing which races it has a read on is real and worth having. It is simply not the same thing as an edge.

That is the honest shape of this experiment, and it is why we publish it as confidently as any winner. The machine is clever enough to know what it does not know. It is not clever enough to make the over-round disappear, because nothing is. If you want to see whether we hold any edge worth staking on, we publish that either way, losses included, in the track record.