The question
"The maths said this losing run was a 1 in 200 event. Why has it happened to me twice in two years?"
Fat tails in betting are the reason. The bell curve from Lesson 1.1 rests on assumptions that real bettors break all the time, and when they do, the extreme results turn up far more often than the curve predicts.
The idea in one sentence
A fat tail means extreme results happen much more often than the bell curve says, and in betting the usual causes are lopsided payouts, linked bets and an edge you don't know as precisely as you think.
The picture
The bell curve makes three quiet promises. Every bet is independent of the others. The true win rate is fixed and known. And there are enough bets for the lumps to smooth out into a neat hill.
Break any one of those and the tails fatten:
- Lopsided payouts. Laying at 12.0 wins a small amount most of the time and loses eleven times the stake occasionally. With a few hundred bets the losses don't smooth into a bell shape; they arrive in lumps.
- Linked bets. Ten lays on the same Saturday share the weather, the referee trends and whatever quirk of the fixture list made the favourites win. They aren't ten independent pieces of evidence.
- An uncertain edge. Your model says the lay loses 7.5% of the time. The truth for this season might be 5%, or 10%. That uncertainty hits every bet at once.
The tool below draws three curves for the same strategy: the bell curve, the exact count for independent bets, and a "clumpy" version where the true rate itself wobbles. Watch the right-hand tail.
| Run | Bell | Exact | Clumpy |
|---|---|---|---|
| 23+ (2 SD) | 1 in 45 | 1 in 36 | 1 in 7.9 |
| 27+ (3 SD) | 1 in 990 | 1 in 448 | 1 in 19 |
| 30+ (4 SD) | 1 in 20,167 | 1 in 4,417 | 1 in 40 |
Worked Betfair example
You lay outsiders in Match Odds at an average of 12.0, with a £10 backer's stake each time. Your liability is £110 per lay. Your model says these outsiders really win 7.5% of the time, against the 8.33% 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%, so £9.80. If it wins you pay £110.
- Expected profit per lay. 0.925 × £9.80 − 0.075 × £110 = £9.065 − £8.25 = +£0.815. Over 200 lays that's +£163 expected.
- Expected losing lays. 200 × 7.5% = 15 outsiders winning. The standard deviation of that count is √(200 × 0.075 × 0.925) ≈ 3.72.
- The disaster scenario. Suppose 25 outsiders win. Your P&L is 175 × £9.80 − 25 × £110 = £1,715 − £2,750 = −£1,035, on a strategy expected to make £163.
- What the bell curve says. 25 is about 2.7 standard deviations above 15. The bell curve gives a chance of 0.54%, about 1 in 186.
- What the exact count says. Counting the actual chances for 200 independent lays (the binomial) gives 0.85%, about 1 in 118. The lopsided payout alone has made the disaster 1.6 times more likely.
- What happens when your 7.5% is uncertain. Suppose the true rate for this batch of lays isn't fixed at 7.5% but averages 7.5% and lands between 5% and 10% about two times in three. Now the chance of 25 or more winners is 8.2%, about 1 in 12.
Verdict: the same "1 in 186" disaster is really closer to 1 in 12 once you admit you don't know your edge exactly. That's the gap between a strategy you'd size aggressively and one you'd size with care.
The same comparison further out
Chance of at least this many outsiders winning in 200 lays:
| Winners | P&L | Bell curve | Exact, fixed 7.5% | True rate uncertain |
|---|---|---|---|---|
| 20 | −£436 | 11.4% | 11.6% | 22.5% |
| 22 | −£676 | 4.0% | 4.6% | 15.4% |
| 25 | −£1,035 | 0.54% | 0.85% | 8.2% |
| 28 | −£1,394 | 0.04% | 0.11% | 4.1% |
| 30 | −£1,634 | 0.005% | 0.02% | 2.5% |
Further out, the gap explodes. The bell curve calls 30 winners a 1 in 20,000 event. With an uncertain rate it's 1 in 40.
The formula
The exact count (binomial)
- K is the number of your selections that go against you (outsiders winning, for a layer).
- n is the number of bets, p the chance each one goes against you.
- (n choose k) counts the ways k results can fall among n bets.
In plain English: this is the true chance of each possible count when bets are independent and p is known. It's lopsided when p is small, which the bell curve ignores.
Clumping: the beta-binomial
- ρ (rho) measures how much the bets move together, from 0 (fully independent) upwards. It's the correlation between any two bets.
- n is the number of bets.
In plain English: even a tiny link between bets balloons the swing when there are many of them. With 200 bets and ρ = 0.01 (the example above), the standard deviation of the count grows by √2.99 ≈ 1.73 times, from 3.72 to 6.44.
Tail ratio
- P real is the chance under the more realistic model.
- P bell curve is what the normal approximation says.
In plain English: how many times more often the disaster really happens. For 25 winners above, 8.2% ÷ 0.54% ≈ 15 times.
Try it
Set lay, odds 12.0, true chance 7.5%, 200 bets and clumping 0.01, and check the "25 or more" line reads about 1 in 186, 1 in 118 and 1 in 12. Then set clumping to zero and switch to backing at 1.4 to see how much tamer a short-priced strategy's tail is.
Common mistakes
- Treating "3 standard deviations" as impossible. For a clumpy, lopsided strategy, 3-standard-deviation losing runs turn up every season or two. Size your bank for them.
- Counting a weekend's lays as independent. Ten bets on one Saturday share conditions. Test your records by weekend or by month, not just bet by bet.
- Trusting your edge to the decimal place. An edge estimated from a few thousand bets has a wide confidence interval. That uncertainty is itself a fat tail.
- Selling the tail without knowing it. Strategies that win small and often (laying outsiders, backing heavy odds-on prices in play) look smooth for months. The bill arrives all at once. Check the worst case, not the average.
- Stress-testing with the bell curve. Use simulation or the exact counts instead (Monte Carlo simulation, extreme value theory). Then read drawdowns and losing runs.
How fat tails catch out good models: why good models still lose money.