Forecasting Correlation, Not Just Volatility

A desk holding long positions in two oil-services stocks feels safe if each one's daily swing looks tame. But if both stocks tend to fall together on the same bad days, the portfolio is really one big position wearing two tickers. Volatility describes how far a single asset moves; correlation describes whether your assets move together, and that second number is what decides whether diversification is actually happening or just an illusion on paper.

Why correlation is the harder number

A volatility estimate uses every return in a single series to answer one question: how big are this asset's moves? A correlation estimate needs enough co-movement between two series to be trustworthy, and true co-movement is a smaller, noisier signal than movement itself. Worse, correlations are not stable through time the way long-run volatility roughly is — they spike toward one during market-wide selloffs and drift apart during calm periods, a pattern sometimes called correlation breakdown.

$$\rho_{ij} = \frac{\text{Cov}(r_i, r_j)}{\sigma_i \, \sigma_j}$$

In words: correlation is the co-movement between two assets' returns, scaled down by how volatile each one is on its own, so it always lands between -1 and +1 regardless of how big either asset's swings are.

Why it moves when you need it not to

Correlations tend to rise exactly when a portfolio can least afford it — during liquidity crunches, macro shocks, or crowded-trade unwinds, most assets get sold for the same reason (raising cash, cutting risk) and their prices start moving together regardless of their usual fundamentals. A risk model built on a calm-period correlation matrix will understate the true risk of a portfolio precisely when a crisis begins, because the diversification it assumed quietly stopped existing.

Drag the correlation slider above from near 0 toward 0.9 and watch the scatter tighten into a line — this is the same picture as two "different" stocks starting to move as one. A portfolio risk model that assumed the loose scatter is silently exposed to the tight one.

Worked example

A fund holds equal-sized positions in Stock A (20% annualized vol) and Stock B (25% annualized vol), estimated at a historical correlation of 0.30. Portfolio variance for two equal weights $w=0.5$ each:

$$\sigma_p^2 = w_i^2\sigma_i^2 + w_j^2\sigma_j^2 + 2w_iw_j\sigma_i\sigma_j\rho_{ij}$$

Plugging in: $0.5^2(0.20)^2 + 0.5^2(0.25)^2 + 2(0.5)(0.5)(0.20)(0.25)(0.30) = 0.01 + 0.015625 + 0.0075 = 0.033125$, so $\sigma_p \approx 18.2\%$.

Now suppose a stress period pushes the realized correlation to 0.85 instead. Recomputing: $0.01 + 0.015625 + 2(0.5)(0.5)(0.20)(0.25)(0.85) = 0.01 + 0.015625 + 0.02125 = 0.046875$, giving $\sigma_p \approx 21.7\%$. The individual assets' volatilities never changed — only the correlation did — and portfolio risk still jumped by nearly a fifth.

What this means in practice

Because pairwise correlations are noisy and there are so many pairs in a large portfolio, most desks do not forecast each correlation independently. They shrink the full correlation matrix toward a simpler structure — a single average correlation, or a factor model where every pair's co-movement runs through a small number of common drivers — because a shrunk, biased-but-stable matrix usually forecasts real portfolio risk better than a noisy, unbiased one.