Install- All races
- Racing info
- Totalizers
- Sandbox
- How it works
- Log in
- Start free account
Join the family, Fat Tony's got your back!
Fat Tony deploys horse and harness scoring together, then automatically routes each race to its specialised model. Both scorers now train separate WIN and PLACE estimates from decision-time prices, runner form, race-relative strengths and price movements.
Specialised CatBoost predictors estimate probabilities. Ask Fat Tony is the conversational AI layer for discussing race information; its written answers are distinct from the numerical model estimates.
One continuous ROI line for the Fat Tony Model bettor, with thoroughbred and harness bets combined. Results carry through model updates without restarting the calculation.
ROI = Fat Tony Model profit ÷ Fat Tony Model stake. Dashed checkpoints mark each version’s first recorded bet, rather than an exact deployment time. Hover or tap a checkpoint for the version.
Evaluation figures will appear here when the active model artifact includes a chronological holdout report.
Fat Tony stores price snapshots as the race approaches and builds comparable views around 60, 30 and 10 minutes before the start, plus the decision-time price at roughly 6 minutes.
Fixed WIN and PLACE odds are converted into implied probabilities and normalised across the field. Each market becomes the starting point for its matching target.
Absolute form is supplemented with within-race differences, z-scores and percentiles for barrier, weight, recent form, ratings, experience and connection statistics.
One CatBoost model learns WIN outcomes and a separate CatBoost model learns actual placings. Each estimates a correction to its corresponding market log-odds using race-relative fundamentals, conditions and earlier price movements.
Corrections are calibrated on later validation races. WIN probabilities total one winner, while PLACE probabilities total the applicable number of paid places. Fat Tony compares each estimate with its matching current price.
Decision-time fixed WIN and PLACE odds, de-vigged market rank and overround, plus WIN price and probability movements from the 60, 30 and 10-minute checkpoints.
Each runner’s barrier, weight, rating, form, career record and connection record are expressed relative to the other runners in this race.
Recent finishes and volatility; career, track, distance and track-distance records; going records; first- and second-up form; time since the last run.
Track, class, surface, weather, direction and going, plus jockey, trainer and recent jockey-trainer results.
V3 does not assume every price move is informed money, and it does not simply follow the favourite. Its two models make separate evidence-based corrections to strong WIN and PLACE market baselines.
| Column | Meaning | How to read it |
|---|---|---|
| Raw model · Win % | V3’s market-anchored and calibrated chance of winning, normalised across the field. | It is the WIN market probability after the learned WIN correction—not a market-free prediction. |
| Raw model · Place % | V3’s separately trained chance that the runner finishes in a paid place. | PLACE probabilities total the number of paid places; they are not derived from the WIN ranking. |
| Raw model · Fair odds | Decimal break-even odds calculated as 1 ÷ the matching V3 probability. | A 20% probability gives fair odds of 5.00. |
| Raw model · Edge | V3 probability minus the matching de-vigged WIN or PLACE market probability, in percentage points. | V3 20%, market 16% = +4.0pp. |
| Market · Win/Place % | Current fixed WIN or PLACE odds converted to field probabilities after removing the bookmaker margin. | This is the crowd’s relative assessment at the scoring time. |
| Decision view · Win % | The probability used by the virtual experiment after operational weighting. | V3 currently runs at 100% model weight, so this normally matches the relevant raw probability. |
| Decision view · EV | Decision probability × fixed odds − 1. | +8% is a model estimate across many similar opportunities, never a promise for one race. |
Fat Tony looks for a model probability at least one percentage point above the de-vigged market. It can still sit out when data coverage is poor, the field is unsuitable, odds are unavailable or outside safety bounds, or no runner clears the configured rules.
Qualifying experimental selections currently use a fixed 10-Tony virtual stake. This deliberately generates forward evidence while keeping results easy to compare. It is not real-money betting and the staking results are not used as proof of model accuracy.
The next useful checkpoint is 300 genuinely unseen races. That is enough to rerun probability and market comparisons with less noise, but it is still an early checkpoint rather than a robust final model.
Check calibration, log loss, Brier score, closing-price movement and at least 30 qualifying virtual bets.
Break performance down by country, odds range, field size, race class, distance and track condition without relying on tiny groups.
Enough evidence to train and test across seasons and common race types. Specialist models will need additional examples inside each specialty.
A sensible target is therefore at least 3,000–5,000 races and roughly 30,000–50,000 runner records, collected across six to twelve months, with the newest 500–1,000 races kept completely untouched for final testing. More data will not guarantee an edge, but it will make it much easier to tell genuine signal from luck.
Version thoroughbred-v3-win-place-2026-10-09-repairs uses 1,240 training races and 11,970 runner records.
Trained through 30 September 2026.