Evidence on Time-Varying Factor Models for Equity Portfolio Construction

Many applicationers derive the variance-covariance matrix (VCM) for mean-variance optimization from some risk model or apply a simple historical estimate. A common problem to these approaches is the stability of the variance-covariance matrix. In turbulent market phases risk estimates from various risk models are well known to be unreliable. One reason for their poor risk forecasting ability is the fact that financial markets are subject to substantial structural change, applied risk models do not account for. In our paper we account for structural changes by deriving VCMs from time-varying estimates of the single factor model, i.e., the market model. We demonstrate the advantages of this approach with respect to risk estimation, portfolio selection and investment performance by means of simulated trading strategies. The problem of choosing the adequate risk model has come in mind of scientific researchers and practioners only recently. While research has focused on forecasting returns for a long time there is a lack of evidence in evaluating the performance of different risk models and the consequences for portfolio optimization. Next to the well known sensitivity of the mean-variance optimization with respect to assumed expected returns the benefits promised by this approach also heavily depend on the accuracy in estimating the VCM (see, for example, [1] and [4]). Given the well known difficulty of estimating expected returns the most important improvement on MV optimization can be made in the VCM estimation which is mainly based on financial econometrics. However, on the performance of alternative risk models and optimization procedures there is only limited scientific evidence, such as [3, 9, 10, 13, 18] among others.