Abstract
Rolling PCA is useful for multi-asset factor analysis but often produces unstable factor spaces that rotate across estimation windows. This paper proposes subspace-regularized PCA, which shrinks the estimated factor space toward an economically motivated prior exposure subspace while preserving rotational invariance. In multi-asset data, moderate regularization reduces mediumhorizon subspace rotation and factor-portfolio turnover with limited loss of reconstruction fit. High-dimensional tests using Fama-French size and book-tomarket sorted portfolios show similar stability-fit-turnover trade-offs. A longonly minimum-variance exercise further shows that the stabilized factor model preserves risk performance while modestly reducing portfolio turnover and drawdown.