Estimating Correlations Between Strategies
Deciding how much diversification a multi-strategy book actually gets requires estimating how correlated its sleeves are, a number that's noisier, more sample-dependent, and more prone to understatement than most allocators expect.
Prerequisites: Defining Sleeves in a Multi-Strategy Book
Allocating capital across sleeves depends entirely on how independent those sleeves actually are. Two sleeves with a correlation of 0.1 diversify each other well; two sleeves with a correlation of 0.7 barely diversify at all, no matter how different their strategy descriptions sound on paper. The number that separates these two worlds — the estimated correlation between sleeve returns — is also one of the hardest numbers in the whole allocation process to get right.
Estimating correlations between strategies means computing this number from historical sleeve returns, and understanding exactly why the naive version of that computation is usually too optimistic.
A correlation estimated from a short or calm history is almost always lower than the correlation that shows up in a real crisis — sleeves that look independent in normal markets often share a hidden common driver that only activates under stress.
Why the naive estimate understates the truth
Correlations are typically estimated from a rolling window of historical daily or weekly sleeve returns — a straightforward calculation. Two problems make the result less reliable than it looks. First, sample size: a two-year window of weekly returns is only about 100 data points, and correlation estimates from that few points have wide error bars, so a measured 0.15 could easily be a true 0.35 by chance. Second, and more important, correlations are not stable through time — many strategies that look uncorrelated in calm markets become correlated in stressed markets, because stress often triggers a shared behavioral response (deleveraging, margin calls, a flight to the same safe assets) that a calm-period sample never captures.
Worked example
A firm estimates the correlation between its stat-arb equity sleeve and its rates relative-value sleeve using two years of weekly returns during a relatively calm market, getting 0.08 — near zero, suggesting strong diversification. Re-running the same calculation using data that includes a 2020-style liquidity crunch, the correlation over that stressed sub-period rises to 0.55, because both sleeves needed to delever into the same illiquid conditions at the same time. Allocating capital as if the true correlation were 0.08 would understate the combined portfolio's stress-period volatility by a wide margin.
What this means in practice
Serious multi-strategy risk management estimates correlation two ways: an unconditional estimate from the full available history, and a conditional estimate computed specifically over past stress periods (or generated by explicit stress scenarios when history doesn't contain enough of them). Sizing the book to the calm-period number alone systematically understates how much the sleeves will move together exactly when losses are largest.
Reporting a single correlation number for a pair of sleeves invites false confidence. The honest version of the number is a range, or better, two numbers — a normal-market estimate and a stress-period estimate — because the gap between them is often the single most important input into how much leverage the combined book can safely carry.
Related concepts
Practice in interviews
Further reading
- Ang, Asset Management: A Systematic Approach to Factor Investing (ch. 3)