Abstract:
Determining a plausible number of components in a factor model is a nontrivial issue in case of weak data, sparse model restrictions and diffuse prior information. We discuss the issue of structural parametric identification in a static factor model and introduce orthogonal restrictions which imply that inference is independent of the order of the dependent variables. Given that financial and economic relations vary over time, we propose the use of predictive likelihoods in combination with moving window estimation in order to determine a plausible time-varying number of factor model components. Results are presented on a residual momentum strategy based on a time-varying latent factor model which outperforms a standard momentum strategy using a portfolio of US industrial stocks.