Two ETFs on related businesses (large banks and regional banks, gold and silver, integrated oil and exploration companies) should not drift apart for long, because the same fundamentals drive both. When the spread between their log prices stretches to an extreme, bet on it closing. This is the textbook relative-value trade studied by Gatev, Goetzmann and Rouwenhorst.
The whole idea rests on one structural fact: that the two prices are anchored to each other, so the spread has a mean to return to.
How it was tested
Pre-registered on 2026-05-23: 10 hand-picked pairs, from GLD/SLV to SPY/RSP, and 12 variants over lookback (60 or 120 days), entry and exit z-scores and a time stop, with a hedge ratio from rolling OLS and equal notional on both legs. Twelve variants put the out-of-sample bar at t > 2.93.
Before building anything, a precondition was registered: each pair had to be cointegrated (Engle-Granger p < 0.05) in at least 50% of rolling windows. Only pairs that passed would be simulated.
Where it died
The precondition check, run the same day, ended it. Zero of 10 pairs passed; the best reached 18%. No strategy backtest was ever run, the pair-aware simulator was never built, and the significance gate was never evaluated. The one pair that looked cointegrated in-sample lost it out of sample.
The record
Thesis
Pairs of ETFs on related underlyings (large vs regional banks, gold vs silver, integrated oil vs E&P) are tied by shared fundamentals, so their log-price spread mean-reverts: enter when the spread's z-score is extreme, exit when it returns to the mean.
Gatev, E., Goetzmann, W. & Rouwenhorst, K. (2006). Pairs Trading: Performance of a Relative-Value Arbitrage Rule. Review of Financial Studies.
Vidyamurthy, G. (2004). Pairs Trading: Quantitative Methods and Analysis. Wiley.
Engle, R. & Granger, C. (1987). Co-integration and Error Correction. Econometrica.
Specification
signal
spread = log(A) − β·log(B), β from rolling OLS; enter when |z| exceeds the entry threshold, exit when |z| falls back below the exit threshold or at a time stop. 12 variants over lookback (60 / 120 days), entry z, exit z and time stop.
No pair passed, so no strategy backtest was ever run.
Raw significance (legacy)not evaluated
Are the strategy's raw returns statistically different from zero?
OOS t > 2.93 (N=12)
Why it died
None of the 10 pairs was cointegrated: the best rolling pass rate was 18% (HYG/LQD), against 50% required. The pairs are correlated day to day — they share factor exposure — but their prices are not anchored to each other, so the spread has no reason to come back. The one pair cointegrated in-sample, EFA/EEM (p 0.008), lost it out-of-sample (p 0.49).
What it taught
Correlation is not cointegration. A spread mean-reverts only when something mechanically ties the two prices together — an ADR and its underlying, an ETF and its futures. Sector ETFs have no such anchor. Testing the precondition first saved building a pair-aware simulator and running twelve backtests.