Bayesian estimation of sparse dynamic factor models with order-independent and ex-post mode identification

Abstract

Common variation in N series is extracted into k≪N dynamic factors. We induce sparsity by using a zero point mass–normal mixture prior distribution on the loadings. Estimation and rotational identification are independent of variable ordering. Sparsity helps identifying the factor space and the factors. Rotational identification, including factor order and sign, is obtained by processing the posterior output and based on factor draws rather than factor loading draws. Simulating data, we document sampler and estimation efficiency. To illustrate, we estimate the model for a large panel of Swiss macroeconomic and detailed price data. We identify 16 factors with a clear economic interpretation.