r_i = α + β₁·MKT + β₂·SMB + β₃·HML + β₄·RMW + β₅·CMA + ε

• α (alpha): the return your portfolio earns after accounting for factor exposure — the holy grail
• βs (factor loadings): how much exposure you have to each systematic factor
• ε (residual): idiosyncratic return — the stock-specific noise

Raw correlation tells you the average relationship. It hides two critical things:

1. Regime-dependent correlations: In a crisis, correlations spike toward 1.0. That "diversifying" asset starts moving in lockstep with everything else. The 2008 crisis, March 2020 COVID crash, and 2022 rate shock all showed correlations converging when you need diversification most.

1. Spurious correlation from shared factor exposure: Two stocks might show 0.3 correlation, but if you decompose them, they're both 80% momentum factor — when momentum reverses, they both get crushed together.

PCA reduces your N×N correlation matrix into a handful of uncorrelated principal components. In practice:

• PC1 almost always maps to "the market" (explains 40-60% of variance)
• PC2 often maps to a growth-vs-value rotation
• PC3+ capture sector rotations, rates sensitivity, etc.

If your first 2 PCs explain 80%+ of your portfolio's variance, you don't have 20 positions — you have 2 bets with a lot of window dressing.

Adding a genuinely uncorrelated return stream to a portfolio improves Sharpe ratio even if the stream itself has a mediocre Sharpe:

Sharpe_portfolio = (w₁·S₁ + w₂·S₂) / √(w₁²+ w₂² + 2·w₁·w₂·ρ·σ₁·σ₂)

When ρ → 0, the denominator shrinks and portfolio Sharpe goes up. This is why hedge funds pay a premium for uncorrelated alpha.