Mean Reversion explained for betting

In one sentence

Mean reversion is the tendency of a price that has moved away from a reference level to move back towards it, and mean-reversion trading bets on that return.

How it works

Betfair prices get pushed around by single large orders. A £2,000 lay can knock a team out several ticks in Match Odds, not because anything has changed about the team but because the order ate through the money waiting at those prices. Often the price then drifts back as other traders refill the gaps.

A mean-reversion trader needs two things: a reference level (a moving average, volume-weighted average price, a model price or a bookmaker consensus) and an estimate of how fast prices return to it. The speed is usually expressed as a half-life: the time it takes for half of a gap to close.

The strategy is the mirror image of momentum. If the move was noise, fading it pays; if the move was information, fading it is standing in front of a train. Most of the skill lies in telling the two apart.

The maths

The standard model is the Ornstein-Uhlenbeck process, which in discrete time becomes:

x_{t+1} - \mu = b\,(x_t - \mu) + \varepsilon_t

\theta = -\ln b, \qquad \text{half-life} = \frac{\ln 2}{\theta}

xₜ: the price at time t, best measured as implied probability.
μ: the reference level the price reverts to.
b: the fraction of the gap left after one step (between 0 and 1; smaller means faster reversion).
θ: the reversion speed per step.
εₜ: random noise.

In plain English: each minute, the gap to fair value shrinks by a fixed fraction, and the half-life tells you how long until it is half gone.

Worked betting example

The away side has been trading around 5.0 (implied 20.0%) in Match Odds for half an hour before kick-off. A large lay order knocks it out to 6.0 (implied 16.7%) with no team news. From past matches you have estimated b = 0.9 per minute for gaps like this (illustrative figure).

Speed. θ = −ln 0.9 ≈ 0.105 per minute. Half-life = 0.693 ÷ 0.105 ≈ 6.6 minutes.
Target. If the pattern holds, in about 6.6 minutes the implied probability should be halfway back: (16.7% + 20.0%) ÷ 2 ≈ 18.3%, a price of about 5.45. You set the exit at 5.5 (ticks are 0.1 between 4 and 6).
If it works. Back £20 at 6.0, then lay £21.82 (20 × 6.0 ÷ 5.5) at 5.5. Profit is £1.82 whatever the result, about £1.78 after 2% commission.
If it does not. Your stop is at 6.6 (ticks are 0.2 above 6). Lay £18.18 (20 × 6.0 ÷ 6.6) for a loss of £1.82 whatever the result.
Break-even. You need 1.82 ÷ (1.78 + 1.82) ≈ 51% of these trades to reach target before stop. The half-life estimate describes the average gap, not every trade, so you need a track record to know if you clear that bar.

Where it's good

Fading moves caused by a single large order in pre-match markets, with no news behind them.
Prices that stray from a strong reference, such as the bookmaker consensus or Betfair SP projections.
In-play overreactions, for example a big drift after a dangerous attack that did not produce a goal.
Linked-market gaps, where one market moves and a related one lags (see pairs trading).

Limitations and pitfalls

Informed money looks like noise at first. A move caused by team news or a gamble will not revert, and fading it is the classic way to lose big.
The reference level itself moves. A 30-minute average is out of date if the market has learnt something new.
Half-lives vary a lot by price band and time to kick-off; one number does not cover all matches.
Betfair's ladder is uneven, so targets and stops that look symmetric in percentage terms are not symmetric in ticks.
Close to kick-off, time runs out: a 6.6-minute half-life is useless with 2 minutes to go.
Commission erodes small edges, which is why the break-even win rate sits a little above 50%.

How to build it

statsmodels OLS or AutoReg to estimate b from gap-to-reference series; pandas for building the references.
Data: time-stamped prices plus traded volume, ideally with order book snapshots to spot single-order moves.
Practical tip: tag each move with whether it came from one large order or many small ones, and estimate reversion separately for each.

Momentum – the opposite bet; the same move can be either.
ARIMA – a negative AR term is short-horizon mean reversion.
Cointegration and pairs trading – mean reversion in the gap between two linked prices.
Moving averages and smoothing – common choices of reference level.
Order book imbalance – helps tell noise moves from informed ones.