Why most 'value bets' aren't: closing odds and market efficiency
Published
Every tips site eventually promises “value bets” — selections where the true probability supposedly beats the bookmaker’s price. We built exactly that feature. Then we backtested it honestly, watched it lose 18% flat-stake, and shelved it. This article is the post-mortem, because the lesson is worth more than the feature.
What “value” actually means
A bet has value when your probability × the offered odds exceeds 1. If you believe a team wins 50% of the time at odds of 2.20, that’s 1.10 — a 10% theoretical edge. Simple arithmetic; the entire difficulty lives in the words your probability.
The experiment we ran
Before launch we walk-forward backtested our Dixon-Coles model over 2,071 matches across three seasons: refit weekly on past data only, predict the following week, compare against the closing odds — the final prices before kick-off. Wherever the model saw a 5%+ edge, we simulated a flat one-unit bet.
The “detector” fired on 87% of priced matches — and returned −17.6%. The model wasn’t finding value; it was finding its own miscalibration. Every gap between our number and the market’s was more likely our error than theirs.
Why closing odds are so hard to beat
The closing price isn’t one bookmaker’s opinion. It’s the equilibrium after thousands of participants — including professional syndicates with injury news, lineup models and decades of data — have pushed the line with real money. Research on betting markets has repeatedly found closing odds to be the best public predictor of football outcomes, and our own numbers agree: de-vigged closing odds scored a Brier of 0.555 in our backtest against our model’s 0.578 (what those numbers mean).
Beating the closing line consistently is the professional standard precisely because almost nothing does it.
The red flags this explains
Armed with that, common tipster patterns become legible:
- “Value bets” on every fixture list. Genuine edges against an efficient market are rare and small. A daily page full of them is a miscalibrated model — or marketing.
- Win rates without probabilities. 60% winners on heavy favourites is worse than chance. Only calibration metrics over a public, complete history mean anything.
- No graded archive. If past picks quietly vanish, the record is fiction. (Ours can’t vanish — the database role that writes predictions has no permission to edit or delete them.)
What would change our mind
We’re not anti-value-bets; we’re anti-unproven value bets. The path back is public: improve the model, calibrate its probabilities against market data, and re-run the same simulation until it clears zero convincingly — at which point the feature ships with the evidence attached. Until then, we show our probabilities next to the market’s and let you see the disagreements yourself.
That’s the deal across all of FootInsights: no claim without a number, no number without a public record.