The question
"If the market is so clever, where does it actually get prices wrong?"
Is the betting market efficient? That question is the right place to start looking for an edge. The honest answer is: mostly yes, especially at the close in big markets. But not everywhere, and not all the time.
The idea in one sentence
A market is efficient when its prices already contain what's known, so no simple rule beats them after commission; the edges that survive live where attention is thin or money is emotional.
The picture
Think of every price as a vote. Thousands of bettors, from casual punters to professional syndicates, put money behind their views until the price settles. On a Premier League Match Odds market an hour before kick-off, that's an enormous crowd, and the price that comes out is very hard to beat.
Efficiency isn't the same everywhere. It varies with how much attention a price gets:
| Where | Why it's efficient, or isn't | What to expect |
|---|---|---|
| Big-league Match Odds at the close | Huge money, sharp bettors, all news in | Very hard to beat |
| Over/Under 2.5 in big leagues | Heavily traded, well modelled | Hard to beat |
| Early prices, days before kick-off | Team news not yet known | Moves a lot; room for good judgement |
| Lower leagues and smaller countries | Less money, fewer sharp models | More room for errors |
| Less-watched markets (First Half Goals, Correct Score) | Priced partly off the main markets | Errors can persist |
| Longshot prices in any market | Money chases big payouts | The favourite-longshot bias |
The favourite-longshot bias
Across many sports and decades of data, one pattern keeps turning up. Longshots win less often than their prices say, and favourites win slightly more. Punters enjoy the chance of a big payout, and they pay for it.
Illustrative Match Odds price bands, with a made-up but typical pattern:
| Price band | Implied chance | Actual win rate (illustrative) | Back ROI after 2% | Lay ROI on liability after 2% |
|---|---|---|---|---|
| 1.40 | 71.43% | 72.5% | +0.9% | −5.1% |
| 2.00 | 50.00% | 49.5% | −2.0% | 0.0% |
| 3.00 | 33.33% | 33.0% | −2.3% | −0.2% |
| 5.00 | 20.00% | 19.0% | −6.5% | +0.9% |
| 10.0 | 10.00% | 8.5% | −16.5% | +1.5% |
| 20.0 | 5.00% | 4.0% | −21.5% | +1.0% |
The backer of longshots loses heavily. The layer of longshots makes a thin margin on a lot of liability. On Betfair the bias is smaller than at bookmakers, because anyone can lay an overpriced outsider, but it doesn't vanish.
Worked Betfair example
You decide to lay outsiders in Match Odds at around 10.0, with a £10 backer's stake each time. That's a liability of £90 per lay. You believe these outsiders really win 8.5% of the time, not the 10% the price implies. Betfair takes 2% commission on net winnings. (Illustrative figures.)
- What each lay pays. If the outsider doesn't win, you keep £10 less 2%: £9.80. If it wins, you pay £90.
- Expected profit per lay. 0.915 × £9.80 − 0.085 × £90 = £8.967 − £7.65 = +£1.317. That's +1.46% on the £90 you risk.
- Over 500 lays. You'd expect 42.5 winners and a profit of about £658.50, having put £45,000 of liability on the line in total.
- If the market is right. With 50 winners (exactly 10%), your P&L is 450 × £9.80 − 50 × £90 = £4,410 − £4,500 = −£90. Commission alone turns a fairly priced lay into a small loser.
- Can you tell which world you're in? The standard error of a 10% win rate over 500 bets is √(0.10 × 0.90 ÷ 500) = 1.34 points. The gap you're betting on is 1.5 points, just over one standard error.
- How many lays to know? For a 1.5-point gap to reach t = 2: (2 × √(0.10 × 0.90) ÷ 0.015)² = 1,600 lays.
Verdict: a genuine favourite-longshot bias of this size is worth having, but 500 lays can't separate it from luck. You need well over a thousand, and your closing line value will tell you sooner whether you're on the right side of the price.
The formula
Bias in a price band
- j is a price band, for example all selections priced 9.0 to 11.0.
- W_j is the number of winners in the band.
- N_j is the number of selections in the band.
- q bar j is the average implied probability in the band.
In plain English: in each price band, does the actual win rate beat or fall short of what the prices said? A negative number means the band is overpriced for backers.
Is the bias real?
- The bottom line is the standard error of the win rate if the prices were right.
In plain English: how many standard errors your band sits away from the market. As in Lesson 1.1, look for t beyond about 2, and stricter if you've tested many bands.
Lay return on liability
- a is the actual win rate of the selections you lay.
- O is the lay price, c the commission (0.02).
In plain English: what you make per £1 of liability. At 10.0 with an 8.5% win rate, that's +1.46%.
Try it
A pen-and-paper check. Over a season you backed 400 favourites at an average of 1.40 and 292 of them won. What was your win rate, your ROI after 2% commission, and is the gap to the implied chance real?
Answer: win rate 73.0% against 71.43% implied. ROI = 0.73 × (1 + 0.40 × 0.98) − 1 = +1.62%. The standard error is √(0.7143 × 0.2857 ÷ 400) = 2.26 points, so t = 1.57 ÷ 2.26 ≈ 0.70. Pleasant, but not evidence of anything yet.
Common mistakes
- Assuming you know better than the close in big markets. Big-league closing prices are very hard to beat. If your model disagrees with them often, the model is usually wrong.
- Finding a bias in 200 bets. Price-band gaps of one or two points need well over a thousand bets to confirm. Check the t-statistic before you believe it.
- Testing dozens of bands and keeping the best. Test 40 bands and a couple will look strong by luck. Correct for it (Lesson 7.4), and check the finding on later seasons you didn't use to find it.
- Backing longshots for the thrill. That's the bias working against you. The average longshot bettor pays the most for the privilege.
- Assuming an edge lasts forever. Markets learn. A bias that's been written about gets traded away. Keep measuring (Lesson 7.3).
Why the market is the real opponent: why good models still lose money.