Odds are a probability in disguise. A 2/1 shot is the bookmaker saying “about a one-in-three chance”. So we checked every price against reality: across 219,521 British runners, how often does each one actually win, and what does backing it return?
| Price | The price implies | Actually wins | Back it, you get |
|---|---|---|---|
| 1/3 | 80% | 78.0% | -3.0% |
| 1/2 | 67% | 66.0% | -1.6% |
| 4/5 | 57% | 56.1% | -2.2% |
| Evens | 50% | 47.1% | -5.6% |
| 6/4 | 42% | 39.6% | -4.6% |
| 2/1 | 33% | 30.8% | -7.0% |
| 3/1 | 26% | 25.1% | -5.0% |
| 9/2 | 21% | 18.4% | -10.9% |
| 6/1 | 16% | 14.0% | -9.6% |
| 8/1 | 12% | 10.0% | -13.1% |
| 12/1 | 8% | 6.6% | -15.7% |
| 20/1 | 5% | 3.9% | -23.3% |
| 33/1 | 3% | 2.6% | -23.7% |
| 50/1+ | 1% | 0.8% | -50.6% |
Every price wins a little less often than it is marked — that gap is the bookmaker's margin. And it widens the bigger the price: an odds-on shot is close to fair, while a 100/1 wins barely one time in three hundred. Short is dependable; long is the dream you pay dearly for.
That table was pure counting. This is the AI doing its own job: reading each horse's form (never its price) to rate its chance. Plot what it rated against what actually happened, across 6,601 real races, and the line climbs true. Its read is about as sharp as the market's, a shade behind, and if anything a little cautious.
That climbing green line is the genuine machine learning at the heart of the Lab: an XGBoost and LightGBM model reading the form. It does not beat the market, but it reads a race about as well.