Autoregressive Poisson (ARP) Affine Term Structure Models

This paper presents tractable discrete time ARP affine term structure models, in which the short rate is driven by autoregressive factors distributed according to Poisson mixtures of various probability densities. ARP models encompass popular ARG and VARG0 affine models based on autoregressive Gamma factors, rule out negative yields and accommodate new homoscedastic terms in the conditional covariance of factors. Evidence from US Treasury yields shows that heteroscedastic ARP, ARG and VARG0 models fit yields similarly and cannot closely match the dynamics of yields volatility. However, simple ARP models with homoscedastic and non-negative factors predict more realistic term premia than ARG and VARG0 models and match the volatility of observed yields better than all heteroscedastic models