The ARCH-LM Test for Volatility Clustering

Anyone who's watched a return chart knows volatility comes in clusters — a few calm weeks, then a burst of large moves, then calm again. This is a well-known stylized fact, but "well known" isn't the same as "confirmed for this specific series right now." Before fitting a GARCH-family model (which exists precisely to capture this clustering), it's worth actually testing whether the clustering is statistically present, rather than assuming it because it's a famous pattern in general. The ARCH-LM (Lagrange Multiplier) test does exactly this: it checks whether today's squared returns are predictable from recent past squared returns, which is the formal definition of volatility clustering.

An analogy: checking if storms come in batches

A meteorologist wants to know if stormy days tend to cluster — is a stormy day more likely to be followed by another stormy day than an ordinary day would predict? One way to check: take the size of each day's weather disturbance (regardless of direction — a hurricane counts the same whether winds blow north or south), and see if today's disturbance size helps predict tomorrow's disturbance size. If big-disturbance days reliably follow big-disturbance days, storms cluster. The ARCH-LM test asks the same question about returns, using squared returns as the stand-in for "size of disturbance," because squaring removes the sign (a +3% day and a -3% day both represent equally large moves) and isolates pure magnitude clustering.

The mechanics, one symbol at a time

Take the residuals $\varepsilon_t$ from a fitted mean model (e.g., after removing any predictable drift), square them to get $\varepsilon_t^2$, and regress today's squared residual on its own recent lags:

$$\varepsilon_t^2 = \alpha_0 + \alpha_1 \varepsilon_{t-1}^2 + \alpha_2 \varepsilon_{t-2}^2 + \cdots + \alpha_q \varepsilon_{t-q}^2 + u_t.$$

In plain English: does knowing how big the last $q$ days' moves were help predict how big today's move is likely to be? If volatility is genuinely clustering, recent large squared residuals should help predict a large squared residual today, making the $\alpha$ coefficients meaningfully nonzero. The test statistic is $LM = n R^2$, where $n$ is the sample size and $R^2$ is the goodness-of-fit of this regression — under the null hypothesis of no ARCH effect (no clustering, $\alpha_1 = \cdots = \alpha_q = 0$), $LM$ follows a chi-squared distribution with $q$ degrees of freedom. In plain English: if the regression explains a meaningful chunk of the variation in squared residuals ($R^2$ well above what pure chance would give with $n$ observations), $LM$ will be large, and a large $LM$ relative to the chi-squared benchmark rejects "no clustering."

Worked example 1: computing LM by hand

Suppose a fitted regression of squared daily returns on their own lag 1 through lag 5 over $n = 500$ observations yields $R^2 = 0.08$. The test statistic:

$$LM = n R^2 = 500 \times 0.08 = 40.$$

The critical value for a chi-squared distribution with 5 degrees of freedom at the 5% level is about 11.07. Since $40 \gg 11.07$, this strongly rejects the null of no ARCH effect — squared returns are clearly predictable from their own recent history, confirming volatility clustering is present and a GARCH-family model is a reasonable next step.

Worked example 2: a series that fails to reject

Now suppose the same test applied to a different, less volatility-clustered series (say, a slowly-trending commodity index with fairly uniform daily swings) gives $R^2 = 0.015$ over $n = 500$ observations:

$$LM = 500 \times 0.015 = 7.5.$$

Since $7.5 < 11.07$, this fails to reject the null — there's no statistically detectable volatility clustering in this series at the 5% level. Fitting a GARCH model here would likely add complexity without capturing any real, exploitable structure; a constant-variance assumption might be perfectly adequate.

What this means in practice

Running the ARCH-LM test before fitting a GARCH model is standard diagnostic practice: it justifies the modeling choice with evidence rather than assumption, and confirms the residuals from a mean-equation fit still have exploitable structure in their variance. It's the natural counterpart to checking residual autocorrelation in the mean itself (see Ljung-Box and Portmanteau Tests) — one test checks the level, the other checks the variance.