Cointegration Breakdown Risk
The biggest risk in a cointegration-based trade isn't a normal spread widening — it's the possibility that the long-run relationship itself stops existing while you're holding the position.
Prerequisites: Cointegration, Error Correction Models
A cointegration test on two years of data tells you the spread between two assets has mean-reverted. It cannot tell you it will keep doing so. Every stat-arb desk that has been burned by a pairs trade has the same story: the historical cointegration test looked fine right up until the relationship it described quietly stopped being true.
Cointegration is a statistical description of the past, not a guarantee about the future. Breakdown risk is the chance that the economic link holding two prices together — shared business, shared risk factor, an index or contract linking them — disappears, and the spread that used to mean-revert simply keeps going.
Why the relationship can break
Cointegration between two series exists because something in the real world ties their prices together: two companies in the same supply chain, a commodity and its refined product, a stock and its ADR, an index and its futures contract. If that underlying link changes — one company gets acquired, a regulatory change severs a business overlap, a contract's terms are amended — the statistical relationship the cointegration test found has no reason to persist, even though nothing in the price history yet shows it. The spread keeps trading like an ordinary asset, wandering wherever fundamentals now push it, and a position built to bet on reversion is left holding a directional bet it never intended to take.
Worked example
A desk trades a pair cointegrated on a rolling two-year Engle-Granger test, entering when the spread is 2 standard deviations wide with a stop at 4 standard deviations, expecting reversion within roughly 15 trading days (the historical half-life). One name receives a takeover approach; its price jumps and decouples from the pair's shared factor. The spread blows through 4 standard deviations and the stop triggers a loss — but crucially, a purely statistical stop assumed the spread was still mean-reverting and merely running wide, when the deal news had actually invalidated the relationship entirely. A desk that re-checks the reason for cointegration, not just the spread level, would have exited or resized on the takeover headline itself, well before the statistical stop fired.
What this means in practice
Managing breakdown risk means treating the economic story behind a pair as a monitored input, not a one-time justification: watching for corporate actions, index changes, and news on either name, and re-running cointegration and structural-break tests on a rolling basis rather than trusting a single historical fit indefinitely.
A widening spread and a broken relationship look identical in real time — both show the spread moving away from its historical mean. The only way to tell them apart is by asking whether the reason the spread used to revert is still true, not by watching the spread itself.
Related concepts
Practice in interviews
Further reading
- Engle & Granger, 'Co-integration and Error Correction: Representation, Estimation, and Testing', Econometrica (1987)