Dynamic Inference in Term Structure Models with Unspanned Latent Risks

Abstract

We propose a parsimonious class of arbitrage-free, yields-only dynamic term structure models (DTSMs) with unspanned latent risks. To enable sequential estimation and forecasting, we develop a Sequential Monte Carlo framework that combines particle learning for static parameters with Kalman filter updates for latent states, yielding joint posterior inference and predictive distributions that account for both parameter and state uncertainty. We use this framework to assess the out-of-sample statistical and economic value of bond return predictability from the perspective of a Bayesian investor. Empirically, we find that unspanned latent factors contain predictive information beyond that embedded in the yield curve, improving out-of-sample forecasting performance relative to standard benchmark models. These gains translate into economically meaningful utility improvements across a range of portfolio settings. Finally, we show that the hidden component of the slope-related risk factor is countercyclical and associated with real economic activity, suggesting that the latent factors capture economically relevant variation not directly reflected in yields.