The correlation trap: why your strategies aren't as diversified as you think.

Most systematic books are run as if the strategies are independent. They are not. Correlation budgeting catches the regime-clustered drawdowns that single-strategy backtesting misses.

How correlated are your strategies, really?

The strategy you built last quarter on EUR mean-reversion is probably 0.4 to 0.7 correlated with the strategy you built six months ago on GBP mean-reversion. Different instruments, different parameter sets, similar structural exposure. When the regime that punishes mean-reversion arrives, both bleed at once.

This is the bit that is difficult to flag in a separate build environment. Backtest each strategy individually, see the equity curve, ship it. Position sizing treats each strategy as a standalone bet sized to its own Kelly fraction or fixed-fractional rule. The portfolio-level variance gets implied, never measured.

What does correlation budgeting look like in practice?

Correlation budgeting helps to overcome a lot of the issues you might find with this.

For each pair of live strategies in your book, calculate the correlation of their daily P&L over the longest live window you have. Below 0.3, treat them as independent. Between 0.3 and 0.6, apply a position-size haircut roughly proportional to the correlation. Above 0.6, you do not have two strategies. You have one strategy in two arrangements.

Above 0.6, you don't have two strategies. You have one strategy in two arrangements.

How does the haircut maths work?

The haircut maths is the simple bit. If your fixed-fractional sizing for a strategy is 0.5 percent per trade and it is running at 0.5 correlation with another live strategy of similar size, the effective combined position is closer to 0.75 percent than to 1.0 percent. The standard move is to halve the fractional sizing on the higher-correlation strategy until the pair behaves as if uncorrelated for risk budgeting purposes.

What does this catch that single-strategy backtesting misses?

Three things.

Regime-clustered drawdowns. When the regime shifts against your dominant structural exposure (carry, momentum, mean-reversion), the correlated strategies all draw down in the same week. A portfolio that looked diversified on paper concentrates at the worst time.

Capacity at scale. Two strategies at 0.6 correlation have roughly two-thirds the effective capacity of two uncorrelated strategies. If you are sizing into allocator capital based on individual-strategy capacity, you are overstating the book's deployable size.

The operational risk of correlated stops. If five strategies hit their drawdown pause within the same fortnight, the book is running on whatever is uncorrelated, which is usually nothing.

How often should you update the correlation matrix?

The correlation matrix update cadence is the practical detail. Monthly is usually enough for a book that does not change composition often. Weekly for a book in active build mode. The matrix should live somewhere the position-sizing code reads from, not somewhere a human eyeballs once a quarter.

Where do most managers get this wrong?

Two ways most managers get this wrong.

First, calculating correlation on returns instead of on P&L. Returns normalise out the position size, which is the thing you are trying to measure. Use P&L in the base currency of the book.

Second, using too short a window. Three months of daily P&L gives you 60 data points, barely enough for one correlation, never mind a matrix of N choose 2. Six months minimum. Twelve months for stability.

The book you actually run is the correlation-weighted portfolio of your strategies. The book you think you are running, if you have not measured the correlations, is the optimistic one. The gap between the two is where the surprise drawdowns come from. For the factor-decomposition side of the same problem, see how stripping out shared vol exposure collapses the diversification claim.