The question was simple and old: can a sensible machine-learning model, built from public form data, beat the starting price? So we built one properly, two algorithms, XGBoost and LightGBM, reading around thirty pre-race signals per runner: the horse's own form, the going, the class, the field, the connections' recent record.
The first result looked brilliant, and it wasn't. Our spring 2026 backtest showed +15.73% on 330 bets, and for a day it felt like the answer. It was an artefact of selection bias: we had swept about 150 filter ideas and reported the shiniest one. On a window the sweep had never touched, the same filter returned roughly −17%. We published that failure rather than the mirage, and it is still the most useful thing this page has taught us.
So we rebuilt it the hard way. Walk-forward testing only (the model never sees a race after the one it is predicting), more data, more features, and isotonic calibration so its probabilities mean what they say. Then the exam that counts: one pre-registered rule, declared in advance, on an untouched two-month holdout: it returned −16.81% on 119 bets. Honest, negative, and exactly the point.
What that established is the foundation everything here stands on: the market is a formidably efficient forecaster, and a well-built model can read races almost as sharply and still lose at the price. Everything below is that finding, kept honest in public, every pick logged before the off, every result scored, updated daily. The full build story · How we test
We keep the model calibrated, and we keep publishing every result, win or lose. The one real edge question, whether its best picks can beat the bookies (not proven, and the market usually wins), runs live in that experiment. And in the Silicon Tipster League it lines up against five AI chatbots on the same races, every day.
Two tests, one honest question
This page is the scientific ledger: every pick the model makes, in every race, stored and scored for how sharp it really is. For the live daily test, its best picks put up against the bookmaker in real time, see Can it beat the bookies? →
One flat £10 on the model's own top pick in every race it has priced, settled to industry starting price with the fallers counted as losses. Irish racing joined the card on 26 July 2026; everything before that date is British. No system, no staking trick, just the model's single best horse, day after day. This is how that balance would have felt in a real wallet.
On 600 race-days the balance finished up on 208 and down on 390. The best single day made +£496; the worst lost £375. Even its longest winning run (8 days) is shorter than its longest losing one (16 days). That is exactly what a market you can't beat looks like.
The running £ balance, every race-day since the record began. Green sections are days it rose, red days it fell. The gold line marks where backtesting ends and live, published-before-the-off predictions begin.
The same record split into calendar months. Green months won, red months lost. A good month happens; a run of them does not.
All-time against the most recent 90 and 30 days, including how the model's probabilities score against the market's on the same races (Brier score, lower is sharper). The market stays ahead.
| Window | Bets | Strike | ROI to SP | Profit / loss | Model Brier |
|---|---|---|---|---|---|
| All time | 9,288 | 24.5% | -13.4% | -£12,439 | 0.0999 |
| Last 90 days | 2,744 | 20.6% | -18.8% | -£5,156 | 0.0938 |
| Last 30 days | 1,154 | 20.4% | -15.8% | -£1,824 | 0.0894 |
Model Brier 0.0999 versus the market's 0.0917 on the same races: the model is well calibrated, just not as sharp as the price. That gap is the whole story, and the reason we show you this instead of selling you tips.
The same ledger, sliced the way the Tipster League slices its board. Pick a timeframe and the graph and numbers re-cut to it. £10 flat on the model's top pick per race, settled at SP.
Profit
−£1,466
ROI
-13.6%
Strike rate
21%
Bets
1,075
Staked
£10,750
Returned
£9,284
This is research, not tips. The model is a calibrated read on a race: it picks winners at a consistent rate and still loses to the starting price, which is the whole honest story. 18+, please bet responsibly.