Everything people ask us

One place to read racing like an analyst. Our in-house model prices every UK race for free, we test every common betting system on real results, five AI chatbots tip live races in a league of their own, and we publish every result, win or lose. It is research, not tips, and everything on the site is reproducible from the conventions we publish.

Start with the Tipster League→

Two algorithms, XGBoost and LightGBM, each get half a vote. They read a horse's own form, the going, the class and the field to set a win probability on every UK runner, jumps and Flat. No tea leaves, no gut feel: the same cold arithmetic on every card, published before the off.

The full story of how we built it→

The classic trap in racing models is leakage, using information that was not knowable before the race. Every form figure the model reads comes only from races run before the day in question, and prices join the data after the off, purely for scoring. One honest nuance: the model is trained on years of market-settled results, so the market's wisdom is baked into what it has learned. That is exactly why it competes in the informed test of our Tipster League rather than the blind one.

How we test, in full→

Its top pick wins 25.4% of its races, and its probabilities are well calibrated: when it says 25%, those horses win about a quarter of the time. On the sharpness score we grade it by (the Brier score, where lower is better) the model sits at 0.102 against the market's 0.093 on the same races — a shade behind the sharpest forecaster on earth, and about twice as sharp as guessing.

What a Brier score measures→The full track record→

Because reading races and beating the price are different jobs. The market's prices already contain nearly everything the model knows, plus the bookmaker's margin on top. So the model can name winners at a healthy rate and still lose at Starting Price — which is precisely what its public record shows, and why we publish it. Watching how close an honest model gets is the experiment.

Watch the live test→

The model's pick and win probability for every UK race today, right beside the market's implied chance, plus a plain read of where the two disagree and which meetings carry the fattest margin. It is the shape of the day's card, told by a machine that has no horse in the race.

Today's card→

Free, every race, no sign-up. The honest catch is the reason it can be free: it does not beat the market, so there is no secret worth selling. It is a calibrated read on each horse's chance, useful for understanding a race — not a tipping service.

We have tested 24 of them on 27,676 real British races, settled at Starting Price with fallers counted as losses, and not one finished in front. Back the favourite and you lose about 8.8% of your stake over time; a random four-fold gives up 62.1%. The bookmaker's margin sits inside every price, and no fixed rule we have tested clears it. The interesting question is which habits cost least — the gaps are big.

Browse the strategy library→The full comparison, in one read→

That is exactly what the What Betting Costs page shows: the same £200 a week through every system, on real results, with the leak per week, month and year and the week-by-week ride each one puts a bankroll through. Percentages hide the pain; pounds do not.

What Betting Costs→

No. We show what the model thinks is fair and where the market disagrees, so you can read a race with better information. Its record at Starting Price is public and it is not in profit. Nothing on this site is a signal to stake, and betting should never be a way to make money. 18+, and BeGambleAware.org is there if it stops being fun.

That is the question the Lab exists to test rather than answer with a slogan. Two live experiments run every day: the model's three best picks at £10 each (a real punter's stake, tracked bet by bet), and five chatbots tipping the same races blind and with the market shown. Both records are public and neither is edited after the fact — watch them build and judge for yourself.

The £30-a-day test→The Silicon Tipster League→

The Lab's two voices. Professor Furlong is the tester: every number he quotes is the real backtest. Pascal is the punter who is sure he has cracked it, and the panel of punters who bring him ideas are openly synthetic characters — modelled archetypes, not real people. The characters make the data readable; the verdicts are always the measurements.

The weekly write-ups→

Results and Starting Prices come from the same public record the industry settles on, covering 27,676 GB races 2023-10 to 2026-06 with a recorded result, flat stakes to Starting Price; non-runners void, fallers and pulled-up settled as losses, joint-favourites split. The model publishes fresh predictions every racing day, results reconcile the same evening, and every figure on the Lab's pages is pulled live from those files as of 12 July 2026 — nothing is typed in by hand, so nothing can quietly go stale.

The Lab runs on exactly this. If there is a system your mate swears by, a bet type you have always wondered about, or a question the board does not answer yet, tell us and the Professor will put it through the same honest settlement as everything else. The best suggestions become experiments, and the write-ups land on the AI Blog.

See what's already been tested→