GARCH-type factor model

A new model is proposed for factor analysis of multivariate time series. The latent factors in the model are linked to observed time series through a deterministic relationship in a manner that is similar to the volatility process of the GARCH model. Mathematically-tractable quasi-likelihood is constructed for the proposed GARCH-type factor model, allowing efficient statistical inference even for high-dimensional time series incorporating non-Gaussian idiosyncratic components. Asymptotic theory for statistical inference of the proposed model is also developed. The applicability of the proposed model to real data is demonstrated through macroeconomic data and forward rate data of bonds. The factors extracted are then utilized to elucidate the risk premium of U.S. government bonds.