Abstract

The Diebold-Li (2006) "Yields-Only" model, extended to the global context by Diebold-Li-Yue (2008), has gained significant popularity in the aftermath of the 2008 Financial Crisis, when regulators placed greater emphasis on the market valuation and accounting of liabilities. Thanks to its parsimony, accurate parameter estimation, and strong forecastability at long horizons, the Diebold-Li model is widely acknowledged as state-of-the-art for yield curve modeling and forecasting. Despite its numerous advantages, the model disregards the in-sample dynamic properties of the yield curve factors, which are crucial for forecasting the yield curve. This thesis aims at developing new data-driven state space models to forecast the co-movement of yield curve drivers of different world regions. The models are designed to preserve the dynamic properties of the yield curve drivers embodied in their underlying data generation processes. In the spirit of Diebold-Li, the models allow forecasting the co-movement of yield curves of different world regions by forecasting their drivers. Using actively traded government bond yields for US and Germany, the modeling approach consists in first conducting a comprehensive study of the dynamic properties of US and German yield curve drivers. This study provides evidence about the stationarity of the US and German slopes, nonstationarity of the levels and curvatures, cointegration structure between the levels and curvatures, and existence of Granger causality among all US and German yield curve drivers. A univariate and multivariate state-space study of outliers and structural breaks reveals alterations in the structure of US and German yield curve drivers in proximity of the 2008 Financial Crisis. These transient changes appear to be synchronized across the drivers and resemble of patches of outliers rather than of structural breaks. A study of the US Fed and the ECB monetary policy predictability allows linking the nature of the outliers to a regime change in the US Fed and ECB monetary policy and an increased ability of market participants in predicting the monetary policy stance after the Financial Crisis. The most blatant outliers can easily be handled in our state-space models with the inclusion of intervention variables in the measurement equation. In a recursive out-of-sample forecasting exercise with the Kalman filter and re-estimation of the parameters every 12 months, we explore the performance of our newly developed state-space models in forecasting jointly the US and German yield curves. The forecasting results are promising, providing evidence that our Full State Space Model (FSSM), accounting for all the dynamic properties of the yield data, outperforms the state-of-the-art Diebold-Li model. In addition, we verify the forecasting power of the curvatures, by developing and forecasting with two additional models for the US and German levels and slopes only, i.e., the FSSMLS and the MShock-FSSMLS. The poor forecasting results at all horizons provide evidence that, for our sample of yields, the curvatures do have predictive power for the US and German yield curves.