Principal Component Analysis and Factor Extraction › Limitations and Practical Considerations

While PCA is a powerful technique, several limitations affect its application to financial data.

PCA assumes that the covariance structure is stable over time. In financial markets, this assumption often fails. Correlations spike during crises, factor loadings drift, and new factors emerge. The principal components extracted from 2005-2007 data looked very different from those needed to explain 2008 crisis dynamics. We address this by using rolling windows for PCA, but this introduces a tradeoff between stability (longer windows) and adaptability (shorter windows).

The yield curve factor structure is relatively stable because it reflects fundamental relationships between maturities. Equity factor structures are more volatile, especially for style factors like momentum or value, which experience periodic reversals in their relationship to underlying assets.

Unlike economic factors (market, size, value), statistical factors from PCA may lack clear interpretation. The second or third principal component of equity returns might capture a blend of industry effects, regional exposures, and macroeconomic sensitivities that resists simple characterization. This makes it difficult to communicate factor exposures to portfolio managers or clients who think in terms of economic factors.

PCA also provides loadings that are only identified up to sign and rotation. The sign of each eigenvector is arbitrary (both $\mathbf{v}$ and $-\mathbf{v}$ are valid eigenvectors), and when eigenvalues are close, the corresponding eigenvectors may mix across different samples.

Accurate estimation of covariance matrices requires observations substantially exceeding the number of variables. The rule of thumb suggests at least 3-5 observations per variable for stable eigenvalue estimates. With 500 stocks and daily data, this requires multiple years of history. For monthly data, the requirements become impractical.

This limitation is particularly severe in the tails of the eigenvalue distribution. The smallest eigenvalues are most prone to estimation error, often being indistinguishable from zero due to noise. Sophisticated techniques like random matrix theory help distinguish signal from noise in the eigenvalue spectrum, but these go beyond standard PCA.

In equity markets, academic research has identified hundreds of factors that predict returns. PCA offers no guidance on which statistical factors correspond to compensated risk factors versus data-mined anomalies. A factor that explains significant variance may offer no expected return premium. Conversely, important economic factors may explain little contemporaneous variance but matter greatly for expected returns. This distinction between risk factors (explaining variance) and priced factors (carrying risk premia) is fundamental but beyond PCA's scope.