Structural Break Tests

A stat-arb desk running a pairs strategy has one recurring nightmare: the historical relationship it's trading was real, and then something changed — a merger, a regulatory shift, a change in index membership — and the spread that used to mean-revert just started drifting. A structural break test is how you check, statistically, whether that has actually happened, rather than guessing from a chart that looks worse lately.

How the test works

The classic version, the Chow test, picks a candidate break date, fits the hedge-ratio regression separately before and after it, and compares the combined error of the two separate fits against the error of one fit spanning the whole sample. If splitting barely helps, the relationship looks stable; if splitting helps a lot, the fit is telling you the coefficients before and after are different enough that pretending they're one relationship is misleading. The test produces an F-statistic: large values mean the improvement from splitting is too big to be explained by chance alone.

Worked example

A pairs desk hedges stock A against stock B with a ratio estimated at 1.20 over the past two years. A sum of squared residuals for that single full-sample fit is 480. Splitting the sample at a date six months ago — where B announced a strategic shift — and fitting separately gives a ratio of 1.35 before and 0.95 after, with combined residuals of 260 across both sub-fits. The big drop in residual error (480 down to 260) from allowing the ratio to differ before and after that date is the signature of a structural break; a spread trade sized on the old 1.20 ratio has been running an increasingly wrong hedge since the announcement.

What this means in practice

Structural break tests are usually run in two modes: confirming a break at a date you already suspect (an earnings date, an index change), or scanning many candidate dates to find the one that maximizes the test statistic. The second mode needs a stricter significance bar, because testing many possible break dates and reporting only the best one is itself a multiple-testing problem.