The Variance Ratio Test

If a stock price is a pure random walk, its returns over any given horizon should be unpredictable from past returns — but that claim has a very specific, testable numerical consequence that goes beyond just "no autocorrelation." Under a pure random walk, the variance of a 2-day return should be exactly twice the variance of a 1-day return, the variance of a 5-day return exactly five times, and so on — variance scales linearly with the horizon. The variance ratio test checks this precise prediction directly, and when it fails, the direction of the failure tells you whether the series is trending (momentum) or mean-reverting.

An analogy: a truly clueless drunk walker versus a stubborn one

Picture a drunk walker taking random steps, each step independent of the last. After 1 minute they've wandered some average distance; after 4 minutes, basic random-walk math says their variance in position (not distance, but its square) should be exactly 4 times what it was after 1 minute — random, uncorrelated steps accumulate variance in direct proportion to time. Now picture a different walker who, after stumbling one way, has some tendency to keep stumbling that way (momentum) — after 4 minutes they've wandered further than variance-scaling would predict, because their steps reinforce each other. And a third walker who tends to correct back toward where they started (mean reversion) ends up wandering less than proportional scaling predicts, because their steps partly cancel. The variance ratio test is exactly this comparison, run on financial returns.

The test, one piece at a time

Define the variance ratio at horizon $k$ as the variance of $k$-period returns divided by $k$ times the variance of 1-period returns:

$$VR(k) = \frac{\text{Var}(r_t^{(k)})}{k \cdot \text{Var}(r_t^{(1)})},$$

where $r_t^{(k)}$ is the $k$-period return (the sum of $k$ consecutive 1-period returns) and $r_t^{(1)}$ is the ordinary 1-period return. Under the pure random walk null hypothesis, $VR(k) = 1$ for every $k$ — variance scales exactly linearly with horizon, as it must if returns are uncorrelated with each other. In plain English: this ratio directly measures whether longer-horizon returns are "too variable" or "not variable enough" relative to what independent, uncorrelated steps would produce. $VR(k) > 1$ means returns show positive autocorrelation (a good day tends to be followed by another good day) — momentum. $VR(k) < 1$ means negative autocorrelation (a good day tends to be followed by a bad one) — mean reversion. The test statistic standardizes $VR(k)-1$ by its estimated standard error under the null and compares it to a normal distribution to judge significance.

Worked example 1: detecting mean reversion in a spread

A stat-arb desk computes the daily variance of a cointegrated pairs spread as $\text{Var}(r^{(1)}) = 0.04$ (in spread-return units squared). The 5-day return variance comes out to $\text{Var}(r^{(5)}) = 0.12$, not $5\times0.04=0.20$ as a random walk would predict. The variance ratio is $VR(5) = 0.12/0.20 = 0.6$, well below 1 — the spread's 5-day moves are only 60% as variable as a random walk would produce, direct evidence of mean reversion, consistent with what a pairs-trading strategy is built to exploit. If the standardized test statistic (built from $VR(5)-1=-0.4$ divided by its standard error) exceeds roughly 1.96 in magnitude, this is statistically significant at the 5% level.

Worked example 2: detecting momentum

A momentum desk checks a small-cap index's daily variance, $\text{Var}(r^{(1)}) = 1.0$, against its 10-day return variance, $\text{Var}(r^{(10)}) = 14.5$, versus a random-walk prediction of $10\times1.0=10.0$. The variance ratio is $VR(10) = 14.5/10.0 = 1.45$ — 10-day returns are 45% more variable than independence would predict, evidence of positive autocorrelation (trending behavior) at that horizon. This is the kind of empirical signature — first documented systematically by Lo and MacKinlay for weekly U.S. equity returns — that underlies momentum strategies: past positive returns statistically tend to be followed by further positive returns over these horizons, at least historically and in this sample.

Compare a pure random walk's spread of outcomes at different horizons against the mean-reverting explorer elsewhere on this site — a variance ratio below 1 looks like the mean-reverting paths staying tighter than a random walk would at the same horizon, and above 1 looks like the paths fanning out faster.

What this means in practice

The variance ratio test is a direct, intuitive way to check market efficiency and to size up mean-reversion or momentum signals before building a strategy on them — a stat-arb desk expects $VR(k)<1$ on candidate spreads, a momentum desk looks for $VR(k)>1$ at the relevant horizon, and both use the test to confirm the effect is statistically real rather than a backtest artifact from a small sample.