Our track record — what the model actually does
Stablebet runs an in-house XGBoost + LightGBM LambdaRank ensemble model with a sentence-transformers commentary RAG layer on UK racing. This page is the running ledger of what it predicts and what actually happens. No cherry-picking, no hidden losses, no “subscribe for picks” paywall.
Our in-house model lost 16.8% ROI on the pre-registered Oct-Nov 2024 backtest window.
This page publishes what it predicts and tracks every result. We do this because nobody else does — the methodology is open, the losses are visible, the analysis is honest. The model output is presented as a comparison to the market, not as a recommendation to back, lay, or stake on any runner.
Read the full methodology in our in-house AI horse-racing model write-up. Track the running ledger on the Stablebet track record page.
Gambling can be addictive. Please bet responsibly. Free, confidential support from GamCare, GamStop and BeGambleAware. See our responsible-gambling page for more.
Our in-house model lost 16.8% ROI on the pre-registered Oct-Nov 2024 backtest window — 119 bets.
This page publishes what the model predicts each day and tracks every result against the actual race outcome. We do this because nobody else in UK racing does — the methodology is open, the losses are visible, the analysis is honest. If a model genuinely worked on UK racing, the people running it would be betting the picks themselves, not selling them.
Read the full methodology and backtest history in our in-house AI horse-racing model write-up. Today's model output is at /model-output/.
How the ledger works
What is logged. Every race the model runs against, every per-runner win probability it outputs, and the corresponding market-implied probability at the time of generation. After racing finishes, the actual finishing position is matched in. Nothing is retro-fitted.
How ROI is measured. Flat-stake win-only on the model's top-rated runner per race at industry SP. We deliberately use the simplest possible reference strategy — the model lives or dies on the basic version, not on a hand-tuned filter.
Each-way is not in the ledger. The model only outputs win probability. Modelling place probability is a different objective with different training data; until that's built, the ledger stays win-only. This is also the cleanest possible test of the question we're asking: does the model beat the market on its primary objective?
Calibration over volume. The headline metric we care about isn't hit rate or ROI in isolation — it's whether the model's predicted probabilities match observed frequencies. A 20%-rated horse should win roughly 20% of the time. The calibration plot below updates as the ledger grows.
Historical backtests
These are the results from before the daily ledger started. They're reproduced from the published methodology write-up in full transparency — including the windows where the model looked profitable on the surface and the pre-registered untouched windows where it didn't.
| Window | Bets | ROI | Status |
|---|---|---|---|
| Feb-Apr 2025 | — | +15.7% | Tuned filter — overfit to window |
| Dec-Jan 2025 | — | -17.0% | Same filter, untouched window — lost |
| Feb-Apr 2024 + Feb-Apr 2025 | ~320 | +27.0% | Re-tuned filter — overfit to combined window |
| Oct-Nov 2024 (pre-registered) | 119 | -16.8% | Authoritative — the headline number |
The Oct-Nov 2024 window was pre-registered in the published methodology before testing. That makes it the only one of these results that isn't vulnerable to the overfitting that every other backtest in the table is. It's the figure we anchor to.
Running ledger
The live ledger lands when the daily cron is wired up.
Once predictions are published daily at /model-output/ and post-race reconciliation runs after each card finishes, this table fills in automatically. Until then, the published methodology write-up contains the historical backtest results that the headline number above is drawn from.
Why we publish this
Most UK racing tipsters never publish a verified track record. The ones that do publish carefully-selected weeks or omit losing months. We publish every prediction and every result because the transparency itself is the point: you don't need to trust us, you can read the ledger. If the model is losing money, it'll be obvious. If it's winning, that'll be obvious too. The alternative — opaque tipsters with selective screenshots — has been the industry default for thirty years.
The model output is research, not a tip service. We don't sell picks. We don't run a subscription. We don't take a cut of any bet you place. What we do run is a free, transparent ledger of one in-house model's attempt to compete with the market — and we're honest about the result whether it works or not.