Most systematic funds over-diversify across weakly orthogonal factors that appear independent in backtests but collapse

Most systematic funds don't blow up because they concentrated too much. They blow up because they diversified into an illusion.

I've watched this pattern repeat for two decades. A fund assembles fifteen, twenty, sometimes thirty factors. Momentum. Value. Carry. Volatility risk premium. Mean reversion across three asset classes. The backtest looks beautiful. The correlation matrix shows low off-diagonal entries. The Sharpe ratio on the combined book looks like free money. Then a real stress event hits, and every single one of those "independent" strategies draws down in the same week. The portfolio that was supposed to be diversified behaves like a single leveraged position. The managers act surprised. They shouldn't be.

The problem is not diversification as a principle. The problem is that the way most quant shops measure diversification is fundamentally wrong. They compute pairwise correlations during normal regimes, average them, and call it a covariance structure. But correlations are not static objects. They are regime-dependent. And the regime where correlation structure matters most, the liquidation regime, is precisely the one where your carefully constructed factor zoo collapses into a single axis: the liquidity axis.

This is the core mechanism that needs to be understood. When capital withdraws from risk assets broadly, the marginal price-setter is no longer the fundamental investor or the statistical arbitrageur. It is the forced seller. Forced selling does not respect your factor taxonomy. A fund redeeming from a multi-strategy platform liquidates its momentum book, its carry book, and its mean reversion book simultaneously. The correlation between those strategies during that liquidation is not 0.15. It is 0.95. And every other fund with a similar structure is doing the same thing at the same time, because the same prime brokers are making the same margin calls using the same VAR models with the same lookback windows.

This is where the standard Markowitz framework fails practitioners, not in its mathematics but in its inputs. Mean-variance optimization assumes a stationary covariance matrix. The entire edifice of Modern Portfolio Theory rests on the idea that historical covariance is a reasonable estimator of future covariance. It is not. Especially not in the tails. Researchers like Andrew Ang and Joe Chen documented this asymmetric correlation phenomenon formally over twenty years ago, showing that correlations between equity factors increase significantly in down markets relative to up markets. The finding was not new even then. Practitioners in 1998 learned it with their balance sheets during LTCM. They learned it again in 2007. And again in March 2020. Yet the architecture of most systematic portfolios still embeds the assumption of regime-stable factor independence.

Let me be more specific about the failure mode. Consider a fund running twenty factors. In a benign environment, many of those factors will show mildly positive returns with low mutual correlation. The combined book looks smooth. But ask a harder question: how many of those twenty factors require functioning liquidity to express their edge? The answer, almost always, is nearly all of them. Momentum requires that you can execute into trends before they mean-revert. Statistical arbitrage requires that convergence trades actually converge, which they don't when counterparties are failing. Carry trades require that the yield differential is not overwhelmed by capital flight. Volatility selling requires that realized vol stays below implied, which it doesn't during a crisis. Strip away the normal-regime veneer and what you have is not twenty independent bets. You have twenty expressions of a single bet: that liquidity remains orderly.

What actually works is less comfortable and less impressive on a pitch deck. You identify two or three strategies whose return-generating mechanisms are genuinely different under stress. Not different in a backtest. Different in their causal structure. One strategy might be long convexity explicitly, profiting from the same dislocation that punishes your core book. Another might operate in a market microstructure that is structurally insulated from institutional flow, perhaps because it trades instruments that levered funds do not hold and therefore do not liquidate. The point is not to find low correlation in sample. The point is to ask: when my primary strategy is losing money, what is the structural reason this other strategy would not also be losing money? If the answer depends on historical correlation rather than on a falsifiable causal mechanism, it is not diversification. It is a backtest artifact.

I have run concentrated books that held through events where peers with far more "diversification" hit their drawdown limits and shut down. The math was not complicated. The discipline was. You have to accept looking unoptimized in normal markets. You have to tolerate a lower Sharpe during calm periods because you are not harvesting every marginal factor premium available. You have to explain to allocators why your backtest does not look as good as the fund running thirty strategies in a zero-rate bull market. Most managers cannot stomach that conversation.

The real question is whether our industry will eventually price regime co-dependency into how we evaluate systematic portfolios, or whether we will keep rewarding the appearance of diversification until the next liquidity event reminds us, again, that twenty correlated bets are still one bet.