Mean-reversion half-life
In an Ornstein–Uhlenbeck process the expected gap from the mean shrinks exponentially. The half-life says how long you should expect to wait for half of it to close.
01The formula
- the value of the spread at time t
- the long-run mean the spread is pulled back to
- the speed of reversion per day
- the size of the daily random noise
- Brownian motion: the random part
- the half-life: the time for the expected gap to halve
02See it move
03Where it is used
- Quant researcherEstimates the half-life of a pairs spread before trading it; too slow and capital is tied up for months.
- Portfolio managerSets holding periods and stop-loss times for mean-reversion trades as a few half-lives.
- Market makerModels how fast inventory and short-term price dislocations decay, to size quote skews.
04Why traders use it
A spread that reverts is only tradable if it reverts fast enough. The half-life turns the abstract speed κ into a number of days you can plan around.
05Worked example
With κ = 0.100 per day, the half-life is 6.9 days. A gap of 3.0 is expected to shrink to 1.50 after one half-life; the simulated paths averaged 1.56 on that day. In discrete daily steps this is an AR(1) with φ = 0.9048.
06How it is built
- Ignore the noise for the average: .
- Solve the equation: .
- Set the gap to half its starting value: , so .
- Sampled once a day the process is an AR(1) with , so you can estimate from a regression of today's spread on yesterday's.
07The common mistake
Expecting each trade to close in one half-life. It is the time for the average gap to halve; with noise, individual paths often overshoot or take several half-lives.
08In Python and LaTeX
import numpy as np kappa = 0.1 print(round(np.log(2) / kappa, 2), round(np.exp(-kappa), 4)) # 6.93 days, phi 0.9048
dX_t = -\kappa\,(X_t - \mu)\,dt + \sigma\,dW_t \;\Rightarrow\; E[X_t] - \mu = (X_0 - \mu)\,e^{-\kappa t}, \qquad t_{1/2} = \frac{\ln 2}{\kappa}09Where it fits
Rather not code the maths yourself? Describe your trading idea and let AI build the bot.
Build a bot with AgenKitThe worked example and the model run the same code, so every number matches. Models simplify markets on purpose. This is education, not financial advice.