Mean Reversion and Statistical Arbitrage › Testing for Mean Reversion

Before trading a mean reversion strategy, you must verify that mean reversion exists in your target series. This verification step is essential. A random walk has no mean reversion: past deviations provide no information about future movements. Trading mean reversion on a random walk is a losing proposition because you are betting on convergence that has no statistical tendency to occur.

The distinction between a mean-reverting process and a random walk can be easy to miss. Both exhibit random fluctuations, and both can appear to drift away from their starting point over short horizons. The difference lies in long-term behavior: a mean-reverting process has a "gravitational center" that pulls it back toward equilibrium, while a random walk has no such anchor and can drift arbitrarily far from any reference point.

The canonical model for a mean-reverting process is the Ornstein-Uhlenbeck (OU) process. This stochastic differential equation captures the essential dynamics of mean reversion in continuous time, and gives a mathematically tractable framework for analysis and parameter estimation. The OU process is defined by:

$$dX_t = \kappa(\mu - X_t)dt + \sigma dW_t$$

where:

• $dX_t$: change in the process value
• $X_t$: value of the process at time $t$
• $\kappa$: speed of mean reversion (higher values mean faster reversion)
• $\mu$: long-run mean level
• $dt$: small time increment
• $\sigma$: volatility of the process
• $dW_t$: standard Brownian motion increment

The intuition behind this equation becomes clear when we examine its two components. The first term, $\kappa(\mu - X_t)dt$, represents the deterministic drift that creates mean reversion. When $X_t > \mu$, the quantity $(\mu - X_t)$ is negative, making the drift term $\kappa(\mu - X_t)$ negative, which pushes the process down toward $\mu$. Conversely, when $X_t < \mu$, the drift is positive, pushing the process up toward $\mu$. The further the process deviates from $\mu$, the stronger this restoring force becomes, much like a spring that pulls harder when stretched further.

The parameter $\kappa$ controls how quickly this reversion occurs. A high $\kappa$ means strong mean reversion, where deviations correct rapidly. A low $\kappa$ means weak mean reversion, where the process meanders more before returning to equilibrium. The second term, $\sigma dW_t$, represents random shocks that continuously buffet the process, creating the fluctuations around the mean that generate trading opportunities.

The half-life of mean reversion provides a practical measure of how quickly deviations decay. It is the expected time for a deviation to shrink by half:

$$\tau_{1/2} = \frac{\ln(2)}{\kappa}$$

where:

• $\tau_{1/2}$: expected time for a deviation to decay by 50%
• $\kappa$: speed of mean reversion

The half-life determines the expected holding period for trades. If the half-life is too long relative to your trading horizon, you may not capture the reversion before other factors dominate. For instance, if a spread has a half-life of six months but you need to close positions within weeks due to margin constraints or reporting requirements, you face significant risk that the spread hasn't reverted by the time you must exit. Conversely, if the half-life is very short, perhaps hours or minutes, the strategy may require high-frequency execution capabilities that introduce additional costs and complexity.

The Augmented Dickey-Fuller (ADF) test is the workhorse for testing whether a time series is stationary (mean-reverting) or contains a unit root (random walk). The test is constructed around a clever regression specification that directly tests for the presence of mean-reverting behavior. The test estimates the regression:

$$\Delta y_t = \alpha + \beta t + \gamma y_{t-1} + \sum_{i=1}^{p} \delta_i \Delta y_{t-i} + \varepsilon_t$$

where:

• $\Delta y_t$: change in the series value at time $t$
• $\alpha$: drift constant
• $\beta$: coefficient on the time trend $t$
• $\gamma$: coefficient testing for mean reversion (unit root)
• $\delta_i$: coefficients on lagged changes capturing serial correlation
• $p$: number of lag terms
• $\varepsilon_t$: error term

The key parameter is $\gamma$, which determines the long-run dynamics of the series. The null hypothesis is that $\gamma = 0$, indicating a unit root: the lagged level $y_{t-1}$ has no power to predict the current change $\Delta y_t$. Under this null, the series follows a random walk and exhibits no mean reversion.

The alternative hypothesis is $\gamma < 0$, indicating mean reversion. The intuition here is subtle but important. A negative coefficient on the lagged level means that when the series is high (large $y_{t-1}$), the expected change is negative (pushing the series down). When the series is low (small $y_{t-1}$), the expected change is positive (pushing the series up). This is precisely the behavior we want: a restoring force that pulls the series back toward its long-run trend.

The lagged changes $\Delta y_{t-i}$ are included to absorb any serial correlation in the data, so the test statistic has the correct distribution. A more negative test statistic provides stronger evidence against the null, indicating more convincing mean reversion.

Once you've established that a series is mean-reverting, you need to estimate the parameters $\kappa$, $\mu$, and $\sigma$ to design trading rules. These parameters have direct practical implications: $\mu$ determines the target level for convergence trades, $\kappa$ determines how long you should expect to hold positions, and $\sigma$ determines the typical magnitude of fluctuations and hence appropriate position sizing.

The discrete-time analog of the OU process sets up parameter estimation:

$$X_t - X_{t-1} = \kappa(\mu - X_{t-1})\Delta t + \sigma\sqrt{\Delta t}\varepsilon_t$$

where:

• $X_t$: value of the process at time $t$
• $\kappa$: speed of mean reversion
• $\mu$: long-run mean
• $\Delta t$: time step size
• $\sigma$: volatility parameter
• $\varepsilon_t$: standard normal random variable

This equation shows that the change in the process depends on how far the current value is from the long-run mean, scaled by the mean reversion speed and time step, plus a random shock. Rearranging terms allows us to formulate a linear regression that can be estimated using standard OLS techniques. We begin by expanding the drift term and regrouping:

$$\begin{aligned} X_t &= X_{t-1} + \kappa\mu\Delta t - \kappa X_{t-1}\Delta t + \sigma\sqrt{\Delta t}\varepsilon_t && \text{(expand drift term)} \\ &= \kappa\mu\Delta t + (1 - \kappa\Delta t)X_{t-1} + \sigma\sqrt{\Delta t}\varepsilon_t && \text{(group terms by } X_{t-1} \text{)} \end{aligned}$$

This algebraic manipulation reveals that the OU process, when discretized, takes the form of a simple autoregressive model. The current value depends linearly on the previous value plus a constant and noise. This matches the form of a simple linear regression:

$$X_t = a + bX_{t-1} + \eta_t$$

where:

• $a = \kappa\mu\Delta t$: regression intercept
• $b = 1 - \kappa\Delta t$: regression slope
• $\eta_t = \sigma\sqrt{\Delta t}\varepsilon_t$: regression residual with variance $\sigma^2\Delta t$

The power of this formulation is that simple OLS regression yields estimates of $a$ and $b$, from which you can recover the underlying OU parameters. Specifically, solving for $\kappa$ from the slope coefficient gives $\kappa = (1-b)/\Delta t$, and substituting into the intercept equation gives $\mu = a/(\kappa \Delta t)$. The volatility parameter can be estimated from the standard deviation of the residuals.