How RaceBot picks each-way value bets
RaceBot doesn't try to tip winners. It looks for each-way bets that are priced wrongly, and that happens most in races where the place terms are generous.
Why each-way terms create value
A bookmaker pays the place part of an each-way bet at a fixed fraction of the win odds, whatever the shape of the race. In some races that formula overpays the place:
| Race | Standard terms | With an extra place | Does RaceBot bet? |
|---|---|---|---|
| Exactly 5 runners | 2 places at 1/4 | 3 places at 1/4 | Yes, both |
| 6 or 7 runners | 2 places at 1/4 | 3 places at 1/4 | Extra places only |
| Exactly 8 runners | 3 places at 1/5 | 4 places at 1/5 | Yes, both |
| 9 or more runners (not 12-15 runner handicaps) | 3 places at 1/5 | 4 places at 1/5 | Extra places only |
| Handicaps, 12 to 15 runners | 3 places at 1/4 | 4 places at 1/4 | No |
| Handicaps, 16 or more runners | 4 places at 1/4 | 5 places at 1/4 | Yes, both |
The model
- Form factors. For every runner, around 200 factors from every UK and Irish result since 2015: speed figures (including best figures on today's going, trip and course), ratings adjusted for weight, official mark against the last winning mark, trainer and jockey records against the market over the last 30 and 90 days, sire records, draw, pace and run style, trouble in running read from race comments, and market support in recent races.
- A machine-learning model (gradient-boosted trees trained to pick the winner within each race) turns those into a strength score. It is retrained every year only on earlier races.
- Model plus market. The score is blended with the bookmaker price, the approach Bill Benter used in Hong Kong. The market knows a great deal, so the model only moves the price where its evidence is strong.
- Place chances. Win chances become exact chances of finishing in each paying place, using finishing-order maths (Plackett-Luce) checked to within a point or two of what actually happened.
- Value. Value is the profit you'd expect per £1 staked if those chances are right. Standard bets need at least 3% over jumps and 6% on the Flat. Extra-place bets are valued by the model and by the prices alone, and bet when the average is at least 10%.
Why every bet comes with reasons
For each bet, RaceBot breaks the model's view into the factors that pushed the horse's chance up or down, then writes them in plain English: for example, proven on the going, a trainer whose runners are beating the market, or a good draw, plus the strongest point against it. It also shows the each-way maths: the place odds and the place chance they're compared with. For many extra-place bets, the main reason is simply that the terms are generous, and the write-up says so.
Staking
Total stake is the bet's value divided by 20 as a share of your bank, never more than 2% on a single bet. Extra-place bets are staked at half that, and no more than 10% of the bank goes on in a day. On a £1,000 bank, a bet with 10% value is £5 in total, £2.50 each way.
How it was tested
Every rule and setting was chosen on 2020 to 2022 only. The rules were then run once on January 2023 to May 2026, settled at starting price on the official results. On that test, with extra-place offers assumed on about 6 eligible races a day:
These are simulated test results, not a live record, and they depend on how many extra-place offers are available. They don't guarantee future returns. The live, proofed record is on the results page, and it is the record that counts.
What was tested and rejected
- Shorter price caps and minimum place chances: cut the profitable bets.
- Laying the win half on the exchange: costs more than it saves, because SP runs well below the exchange price.
- Blending in exchange prices: worked on Irish races, failed an independent GB test.
- Extra places on 12 to 15 runner handicaps: lost money in 2020 to 2022.