A Time-Varying Factor Approach to Covariance Estimation

Financial volatility changes continuously, yet most portfolio optimization strategies rely on static covariance estimates. This thesis introduces a time-dependent covariance matrix that updates for each time period via kernel-weighted local principal components which is applied in a time-varying minimum-variance portfolio (TV-MVP). The full workflow is implemented in the open-source R package TVMVP. The package also includes a test for constant factor loadings, evaluated through Monte Carlo simulations across six data-generating processes. Results confirm its statistical reliability: the test accepts constant loadings in over 80% of runs and consistently rejects time-varying structures. Empirically, TV-MVP is applied to 50-250 Swedish stocks and benchmarked against sample, Ledoit-Wolf, EWMA, POET, and graphical-lasso covariance estimates. In calm markets (2017-2019), TV-MVP achieves competitive risk levels and returns. In the volatile 2022-2024 window, it delivers the lowest or second-lowest standard deviation across all asset pools while preserving competitive Sharpe ratios and drawdowns. TV-MVP therefore offers a practical, low-volatility alternative with accessible implementation through its accompanying software package.