How do you handle highly collinear features in a high-frequency financial dataset without losing predictive power?
