Guides / Correlation Between Trading Strategies, Explained

Two strategies · one account

Correlation between strategies, what the number actually hides.

A low correlation number across a full backtest can still mean two strategies lose money on the same days. Here's what correlation actually measures, where the full-period number misleads, and a worked example.

Cephic TeamUpdated Nov 2, 20267 min read

Quick answer: correlation measures how much two strategies' returns move together, from −1 (opposite) to +1 (identical). Measured across a full multi-year backtest, it's calculated from daily or per-trade returns across every day both strategies traded, including the long calm stretches where neither is doing much. That full-period number can look low even when the two strategies behave almost identically during the handful of weeks that actually matter: their worst ones.

WHAT CORRELATION ACTUALLY MEASURES

What the number is actually measuring

For every day both strategies traded, correlation checks whether they tended to be green together, red together, or showed no pattern between them at all. A reading near +1 means they're almost always green or red on the same days. A reading near −1 means one tends to be up when the other is down. A reading near 0 means there's no real relationship between the two.

Most strategy pairs land somewhere between 0.1 and 0.5: loosely related, which is normal and not automatically a problem. The issue isn't the number itself. It's which days are driving it.

The window you measure over also matters. Correlation calculated from daily returns over a three-year backtest and correlation calculated from the same two strategies' individual trades over the same period can come out meaningfully different, because daily returns smooth over the exact sequencing within a day and per-trade correlation doesn't. Neither is "wrong"; they're answering slightly different questions, and it's worth knowing which one a tool is actually giving you before you size against it.

THE FULL-PERIOD BLIND SPOT

Why the full-backtest number can hide the scenario you care about

Take a trend-following strategy on EURUSD and a momentum breakout strategy on GBPUSD. Across three years of daily returns, full-period correlation comes out to 0.18, genuinely low, because on most days the two are unrelated. Now isolate just the dates inside each strategy's five worst drawdown stretches, the weeks that actually threaten the account, and measure correlation only across those overlapping bad weeks. It rises to roughly 0.6.

The two strategies don't resemble each other most of the time. But a shared trigger (a surprise rate decision, a sudden trend reversal across major pairs) hits both of their underlying logic at once, in the exact weeks where it costs the most. A full-period number of 0.18 and a bad-weeks number of 0.6 describe two different things. The first tells you how the strategies behave on an average day. The second tells you how they behave on your worst day, which is the one a position-sizing decision actually needs to survive.

WHAT THIS CHANGES ABOUT SIZING

What this changes about sizing

Strategy A is backtested to a 9% max drawdown at full size, run alone. Strategy B, also 9% at full size, run alone. A naive sum if both run together would put combined worst-case drawdown at 18%. If the two were genuinely uncorrelated during their drawdown periods, not just on average, realistic combined drawdown might land closer to 11 to 12%, since the bad stretches rarely stack on the exact same days. But with a bad-weeks correlation closer to 0.6, like the example above, combined drawdown can climb back up toward 15 to 16%, worse than the full-period number would suggest, though still better than a flat 18% sum.

That gap, between what the full-period number implies and what the drawdown-period number actually shows, is the entire reason to check correlation during the bad stretches specifically, not just the headline figure from the full backtest.

It's also worth being specific about what's actually shifting here. The two strategies' individual edges haven't changed: each one still has the same backtested win rate and average loss it always had. What changes is the timing overlap: how often their bad trades land on the same day instead of spread across different ones. Position sizing on its own can shrink how much any single strategy loses on its worst day, but it can't do anything about two strategies choosing the same worst day in the first place. Only genuinely low drawdown-period correlation does that.

COMMON MISTAKES

Where this gets misread

WHAT TO DO WITH IT

What to actually do with a correlation number

Recompute it whenever you add a new strategy or after live performance drifts meaningfully from the backtest. Look specifically at correlation during each strategy's drawdown periods, not just the full-period average. That's the number that actually matters when things go wrong at the same time. Treat correlation as one input to a weight split, not a single pass-or-fail gate, and keep in mind it describes the past. Like win rate, it can start drifting before the sizing built around it catches up.

Allocate computes correlation across your actual uploaded backtests, including the drawdown-period view, instead of a single full-period guess.

Open Allocate