We tested every famous horse racing betting edge. Here's what actually survived.

Every tipster site has a number. +15% ROI. +27% since launch. 9 winning months in a row. We had one too — our model's +8% backtested ROI was on this very site until this week.

This is the story of how we found out that number was wrong, what we did about it, and what a week of brute-force testing — 13½ years of data, over 600 strategy variants, statistics specifically designed to catch self-deception — says actually works in UK and Irish racing. Spoiler: almost nothing, and the exceptions are not what the adverts claim.

Why you should trust a site that just admitted its number was wrong

Because the alternative is trusting sites that never check. Here's the uncomfortable mechanic of this industry: impressive backtest numbers are easy to manufacture by accident, and almost nobody looks for the accidents. We found eight of them in our own pipeline in one week. Each one made our results look better than reality. Any one of them, left unfound, becomes a marketing headline.

The eight ways our own system lied to us

1. The non-finishers were invisible. Horses that fell or pulled up — about 8–9% of jumps runners, every one a losing bet — were silently missing from our results data. Removing only losers flatters everything.
2. Trainer "14-day form" contained the future. The data feed recomputes trainer form on older racecards in a way that includes results from after the race. Our model looked brilliant on old data because it was reading tomorrow's newspaper.
3. Old odds were too good to be true. Historical racecards carry price lists where the "best price" systematically exceeds what was actually available. Backtesting against the best of 28 books made everything profitable; the median book told the truth.
4. The evaluation quietly preferred easy races. A filter requiring every runner to have complete price data dropped big competitive fields — inflating blind "back everything" returns from roughly −2% to +11% before we caught it.
5. A "certified" each-way edge needed a crystal ball. Our most exciting result — +29% on each-way bets, validated across sealed holdout data — turned out to be selecting bets using the closing price, which doesn't exist when a morning tip goes out. Every version a real punter could actually bet lost between 25% and 50%.
6. Survivorship in the strategy search. Run 500 strategy variants and the best handful look amazing by pure luck. We used multiplicity-corrected statistics (the same maths quant funds use) — and they killed every "discovery" our search produced.
7. Lucky eras masquerade as skill. Strategies that printed +20% in 2016–2019 decayed to +2–5% by 2025–26 as markets got sharper. A pooled "ten-year backtest" hides that the edge is already mostly gone.
8. A one-line code bug made four different strategies report identical results — caught only because we read the output suspiciously. Bugs flatter backtests far more often than they hurt them, because the flattering ones get believed and the ugly ones get debugged.

What actually survived

After all eight were fixed, here is the complete honest scorecard:

• Beating the Betfair closing market with form data: no. Our best model — 87 features, three years of data, properly validated — adds essentially nothing to the closing price. The market at the off is the best forecaster in racing. Anyone claiming to out-predict it consistently should publish a live ledger.
• Our own old +8% claim: corrected. Re-tested leak-free over three years, the strategy made about +5% before commission at exchange prices (roughly breakeven to +3.6% after commission) — and was negative at bookmaker starting price. The corrected numbers now appear everywhere on this site.
• Morning prices are genuinely soft — but the edge is small and shrinking. Predicting which morning prices will shorten works (we measured it across eleven straight years), but the profit decayed from double digits in 2016 to roughly +2–3% gross by 2026, and morning markets are too thin to bet big. Useful, real, modest.
• Each-way "free money", lay systems, exchange pattern-trading: dead at bettable prices. Every version that doesn't require knowing the future lost money after commission in 2025–26 data.
• One genuine curiosity remains open: a pocket of jumps-racing longshots our model flags kept making ~+20% even in data it never trained on — but at ~10 picks a month, statistics can't yet separate skill from luck. We're paper-tracking it live rather than selling it. If it's real, you'll see it prove itself on the ledger; if not, you'll see that too.

What this means if you bet

1. Treat every ROI headline without a live, unedited ledger as fiction. We just demonstrated how easily those numbers manufacture themselves.
2. Closing Line Value is the only honest scoreboard. If a tipster's advised prices consistently beat the closing price, the record is real — it's the one metric you cannot cherry-pick after the fact. We publish ours on every pick.
3. The realistic ceiling is small. Low single digits at exchange prices, sustained, is genuinely good. Everything above that, demand extraordinary evidence.
4. The real wins for an everyday punter are practical: better prices (knowing when to take a morning price before it shortens), faster form study, honest accounting of your own betting. That's what we build.

Where Racing Alpha goes from here

Our ratings, the full daily card, the public ledger and the CLV record stay — free, logged before the off, settled in public, with this audit as the standing proof that when our numbers are wrong, we say so. We'd rather be the only site you don't have to take on faith.

All testing details — the leak audits, the statistical gates, the strategy sweeps — are versioned in our research notes. 18+ only. Betting involves losing money; nothing here is financial advice. If gambling is a problem for you, free confidential help is available at BeGambleAware.