Quant Analytics: from scratch to pro · Factors, volatility, regimes and lead-lag · Pro
Cross-asset lead-lag, cointegration and spreads
Some markets seem to lead others, and some pairs of prices stay tied together. Both ideas are easy to see and easy to fake: non-synchronous closes create apparent lead-lag, and trending prices create apparent cointegration. This lesson shows how to measure each, and what a spread's half-life says.
In this lesson
- Measure lead-lag with cross-correlation and judge whether it is statistically real
- Explain how non-synchronous closes and time zones create false lead-lag
- Describe the Engle-Granger two-step test and what cointegration means
- Compute the half-life and the standard deviation of a mean-reverting spread
Read first: Time series basics: random walks, autocorrelation and stationarity, Correlation and its limits, Volatility: from daily moves to an annual figure
Do some markets lead others? Do some pairs of prices move together over the long run, so that a gap between them tends to close? Both ideas are old, and both are easy to find in a chart and hard to confirm. This lesson separates the measurement from the story.
Lead-lag and cross-correlation
The cross-correlation at lag k is the correlation between market A's return on day t - k and market B's return on day t. A value clearly different from zero at k = 1 says A's yesterday relates to B's today. See lead-lag.
Judge the size against its noise. The standard error of a correlation is about 1 / sqrt(n): 0.063 with 250 days, 0.032 with 1,000. A lag correlation of 0.15 over 1,000 days has t = 4.74, so it is unlikely to be chance. It still explains only 2.25% of the variance. Reaching t = 2 for a correlation of 0.10 needs about 400 observations, and scanning many lags and pairs creates false hits (see multiple testing).
Non-synchronous closes
A market in Asia closes hours before one in the United States. News that arrives in between is reflected in the American close but, for the Asian market, only in its next day's return. A one-day lag correlation then appears even though no market "led": both reacted to the same news at different times. Align timestamps before looking for leadership, compare returns over the same clock interval, and take care with time zones and daylight saving changes. See clean data.
Cointegration
Two price series that each wander (they have unit roots) can still be tied: some combination of them is stationary. Engle and Granger (1987) called this cointegration. If A and B are prices, the combination is a spread:
spread_t = ln A_t - hedge x ln B_t
The Engle-Granger two-step method first estimates the hedge ratio by regression, then tests whether the residual spread has a unit root. The critical values are not the standard ones, because the hedge ratio was estimated first. See cointegration and, for the pitfall, spurious regression (Granger and Newbold, 1974).
A regression slope from returns with correlation 0.9 and volatilities of 1.2% and 1.0% gives a hedge ratio of 0.9 x 1.2 / 1.0, which is 1.08. (This is a hedge for daily moves; the Engle-Granger step regresses the levels, and the two ratios need not agree.) Cointegration is about the levels, not the daily returns; high return correlation is neither necessary nor enough.
Worked example: the spread's half-life
Suppose the spread follows an AR(1): spread_t = 0.95 x spread_(t-1) + e_t, with daily noise of 0.5%.
- Half-life of a gap:
ln(0.5) / ln(0.95), which is 13.5 days. - Stationary standard deviation:
0.5% / sqrt(1 - 0.95^2), which is 1.60%. - A two-standard-deviation gap is 3.20%. If it closes fully and round-trip costs are 0.30%, the gain is 2.9%.
The chart shows the expected gap fading. A coefficient of 0.98 has a half-life of about 34.3 days, and the same gap takes more than twice as long to pay. The half-life sets how many trades fit in a year and how long capital waits; a half-life near the length of the sample means the spread is hardly estimable.
The evidence on pairs
Gatev, Goetzmann and Rouwenhorst (2006) tested a distance-based pairs rule on daily US stock data from 1962 to 2002. They reported average annualised excess returns of up to 11% for self-financing portfolios of pairs, above conservative transaction cost estimates. A published result of this kind is a record of one sample and period. It does not show that the gap will close next time.
Common mistakes
- Reading a one-day lag correlation as leadership without aligning the clocks.
- Calling two trending series "related" after a regression on levels.
- Treating the hedge ratio and the mean of the spread as constants. Both drift, so re-estimate on a rolling window and test it out of sample.
- Ignoring costs on both legs, borrowing costs on the short leg and the risk that the spread widens before it closes.
- Searching thousands of pairs and trading the best. Count every pair tried.
- Using the whole sample to choose pairs, then testing on that same sample.
In practice
Keep three things apart: the statistical tie (a test on the spread), the economic reason for it (same business, same cash flows, same underlying), and the cost of exploiting it. Without the second, the first is a pattern waiting to break. Analytics on FinUties applies measures like these to the markets it covers, so they are at hand when you read a chart.
Expected gap remaining as a spread reverts
- AR(1) coefficient 0.95
- AR(1) coefficient 0.98
Worked numbers
Each result is recomputed from the formula shown.
| What | Formula | Result |
|---|---|---|
| standard error of a correlation, 250 observations | 1 / sqrt(250) | 0.063 |
| standard error, 1,000 observations | 1 / sqrt(1000) | 0.032 |
| t-statistic of a lag correlation of 0.15 | 0.15 / se1000 | 4.74 |
| variance explained | 0.15**2 | 2.25% |
| observations to detect a correlation of 0.10 | (2 / 0.10)**2 | 400 |
| hedge ratio from correlation 0.9 and volatilities 1.2% and 1.0% | 0.9 * 1.2 / 1.0 | 1.08 |
| half-life of the spread, days | ln(0.5) / ln(0.95) | 13.5 |
| half-life at coefficient 0.98, days | ln(0.5) / ln(0.98) | 34.3 |
| stationary standard deviation of the spread | 0.005 / sqrt(1 - 0.95**2) | 1.60% |
| a two-standard-deviation gap | 2 * sd_spread | 3.20% |
| gap after round-trip costs of 0.30% | gap2 - 0.003 | 2.9% |
Check yourself
The correlation between yesterday's return in market A and today's return in market B is 0.15 over 1,000 days. Is it distinguishable from zero?
The standard error is about 1 / sqrt(1000) = 0.032, so t is about 4.7. Statistically yes. It explains only 0.15^2 = 2.25% of the variance, and it may reflect different closing times, not a real lead.
Why can a market that closes earlier appear to lead one that closes later?
News arriving after the first close shows up in the first market's next return and in the second market's current return. A lag of one bar then looks like leadership, though both reacted to the same news.
Two prices each trend upward. Their regression has a high R-squared. Are they cointegrated?
Not necessarily. Two independent trending series can show a high R-squared (spurious regression). Cointegration needs the residual spread to be stationary, which is tested with unit-root tests.
A spread follows an AR(1) with coefficient 0.95 a day. What is its half-life?
ln(0.5) / ln(0.95) = 13.5 days.
Why can a pair that was cointegrated in the past stop being so?
The tie is a statistical property of a sample. Fundamentals change, and the relationship can break; the estimated hedge ratio also drifts.
Key terms
Checked facts
Every figure below carries how it was checked: Primary checked (read in the cited primary, official, academic or data source), Two sources agree, or Approximate (an order of magnitude only).
- Primary checked
Gatev, Goetzmann and Rouwenhorst (2006) tested pairs trading on daily US stock data from 1962 to 2002 and reported average annualised excess returns of up to 11% for self-financing portfolios of pairs, above conservative transaction cost estimates.
Sources
- Co-Integration and Error Correction: Representation, Estimation, and Testing. Econometrica 55(2), 251-276 (Robert F. Engle, C. W. J. Granger), 1987. Accessed 9 Nov 2026
- Pairs Trading: Performance of a Relative-Value Arbitrage Rule. The Review of Financial Studies 19(3), 797-827 (Evan Gatev, William N. Goetzmann, K. Geert Rouwenhorst), 2006. Accessed 9 Nov 2026
- Spurious Regressions in Econometrics. Journal of Econometrics 2(2), 111-120 (C. W. J. Granger, P. Newbold), 1974. Accessed 9 Nov 2026
- Distribution of the Estimators for Autoregressive Time Series with a Unit Root. Journal of the American Statistical Association 74(366), 427-431 (David A. Dickey, Wayne A. Fuller), 1979. Accessed 9 Nov 2026