Updated
AI horse racing accuracy depends entirely on how you measure it. Across UK and Irish racing, a competent AI rating model picks the winner around 18-25% of the time (one in four-to-five races), the top-3 finisher around 50% of the time, and produces a positive return on investment between +5% and +12% on level stakes. Racing Alpha's v1.0 picker backtests at roughly +8% ROI; the v1.1 ML model at +10.3%. These are honest numbers — AI doesn't pick every winner, and any tipster claiming otherwise is selling you noise.
Around 18-25% for a model picking one horse per race across UK & IRE. Above 25% sustained over hundreds of picks is exceptional; above 30% is implausibly good (almost certainly small-sample fluke or curated sample). Below 15% suggests the model isn't beating the favourite. The strike rate alone tells you very little without average price — a 20% strike at 6.0 average odds is profitable, a 30% strike at 3.0 is breakeven.
Anything sustainably above 0% is genuinely rare. +3% to +6% over 500+ picks is solid; +8% to +12% is excellent. Negative ROI doesn't mean the model is bad — it means it isn't currently beating the bookmaker's margin (typically 15-20% built into UK book SPs). Racing Alpha's published v1.0 model targets +8% ROI; v1.1 targets +10%. We publish the live ledger so the numbers can be checked, not taken on trust.
The market favourite wins roughly 32-35% of races at an average SP of 2.5-3.0. A pure favourite strategy returns -10% to -15% long term (the bookmaker margin). A good AI model picks the favourite less than half the time — its edge is finding non-favourites that the market has underpriced. Racing Alpha's v1.0 advises an average price around 5.0-6.0, well above the favourite.
Three honest reasons. One: racing is high-variance — even a strong favourite loses 65% of the time. Two: AI sees the data the market has already priced in; the market itself is wise to obvious signals. Three: every race has unmodellable factors (horse mood on the day, jockey error, in-running interference) that no rating model can capture. Long-term edge isn't about picking every winner — it's about being right slightly more often than your prices imply.
Three checks. (a) Are picks published BEFORE the off, with a stable URL? Anything post-hoc is worthless. (b) Is there a CLV (Closing Line Value) figure published? Positive CLV is the only credibility metric markets respect. (c) Is the sample size 100+ settled picks? Anything under 50 is noise. Racing Alpha publishes all three: live tips, CLV ledger, and 5,000+ race backtests.
Racing Alpha is for information and entertainment for adults aged 18 and over. We do not accept bets. Past performance does not predict future results. BeGambleAware.org