Skip to content
Quantoros

How to deal with the multicollinearity problem of factors?

Quant ResearchLevel: Intermediate

How to deal with the multicollinearity problem of factors?

Hints
  • Identify the financial object, available information and source of risk.
  • Separate theory from empirical validation with costs and constraints.

Answer

Highly correlated factors can make estimated coefficients unstable without necessarily destroying prediction. Inspect correlations, condition numbers and exposure redundancy, then compare grouping related signals, ridge/elastic-net regularization, PCA or residualization. Orthogonalization depends on ordering and PCA components can lose economic interpretability. Fit every transformation within training windows and judge the combined portfolio out of sample, including turnover. A VIF threshold is a diagnostic convention, not an automatic deletion rule.

Source and licence: Quant Research Interview Notes — SoYuCry; English translation by photon2563 · MIT · answer: Quantoros editorial

Related questions