BTC and ETH share a driver, the crypto market itself, and their prices are cointegrated: the spread between them is anchored. When it stretches beyond 1.5 standard deviations, the bet is that it snaps back within 5-20 days. Buy the laggard, short the leader, close on convergence. The position is market-neutral; what it trades is the relationship.
statistical arbitrage / cointegration
BTC/ETH pairs trading
The idea
How it was tested
Pre-registered on 2026-07-31. The signal: a valid cointegration test and a spread z-score above 1.5. Exit when the z-score falls below 0.3, after 20 days, or at a 3% loss per leg. One pair, 2x margin on perpetuals. Costs were assumed at 0.05% per leg per side plus funding, about 0.45% per 15-day cycle. The ficha's own gates asked for at least 10% a year in-sample and a win rate of at least 50%.
Where it died
It was registered on 2026-07-31 already marked validated, with 134.2% a year in-sample and 120% out-of-sample. That rested on 6 in-sample trades and 2 out-of-sample.
The backtest four days later, on 2026-08-04, gave an in-sample return of −571% with a 32% win rate, and 186% out-of-sample. Cointegration itself held, at a Johansen p below 0.001. The relationship was real; the reversion did not arrive on any schedule a 20-day trade with a 3% stop could use. It was archived the same day.
The record
Thesis
BTC and ETH share a fundamental driver and are cointegrated, so when their spread stretches beyond 1.5 standard deviations it should revert within 5-20 days. Long the laggard, short the leader, close on convergence.
Specification
- signal
- cointegration valid and spread z-score > 1.5
- exit
- z-score < 0.3, 20 days, or 3% loss per leg
- sizing
- 2x margin on perpetuals, one pair
- costs assumed
- 0.05% per leg per side plus funding, 0.45% per 15-day cycle
Gates
Pre-registrationpass
Was the hypothesis, its parameters and its pass/fail rule written down before any data was looked at?
Written before the run, but the "validated" ficha already quoted a result (6 in-sample trades) as if it were a pass.
Raw significance (legacy)fail
Are the strategy's raw returns statistically different from zero?
in-sample return ≥ 10% annualized and win rate ≥ 50% (the ficha's own gates)
is return pct-571oos return pct186win rate pct32
Raw returns only — no benchmark, no t-stat. Cointegration itself held (Johansen p < 0.001).
Why it died
The "validated" result that registered it was 134.2% a year in-sample and 120% out-of-sample — on 6 and 2 trades. Re-run four days later, the in-sample figure was −571% with a 32% win rate. The relationship was statistically real (cointegration held) but the spread did not revert on any schedule a 20-day trade could use. The −571% cannot be a compounded annual return at 2x leverage, so the harness was crude too; nothing here is a clean measurement, but there is no sign of an edge to rescue.
What it taught
Cointegration is a statement about where a spread ends up, not when. A pair can be cointegrated at p < 0.001 and still lose money for a trader with a 20-day horizon and a 3% stop. And an annual return computed from 6 trades is an anecdote, not a pass.
Verdict history
- 2026-07-31Registered as validated: 134.2% / 120% annualized on 6 + 2 trades
- 2026-08-04Backtest: in-sample −571%, out-of-sample 186%, win rate 32%; archived No alpha