Combining Canadian Interest-Rate Forecasts

Model risk is a constant danger for financial economists using interest-rate forecasts for the purposes of monetary policy analysis, portfolio allocations, or risk-management decisions. Use of multiple models does not necessarily solve the problem as it greatly increases the work required and still leaves the question "which model forecast should one use?" Simply put, structural shifts or regime changes (not to mention possible model misspecifications) make it difficult for any single model to capture all trends in the data and to dominate all alternative approaches. To address this issue, we examine various techniques for combining or averaging alternative models in the context of forecasting the Canadian term structure of interest rates using both yield and macroeconomic data. Following Bolder and Liu (2007), we study alternative implementations of four empirical term structure models: this includes the Diebold and Li (2003) approach and three associated generalizations. The analysis is performed using more than 400 months of data ranging from January 1973 to July 2007. We examine a number of model-averaging schemes in both frequentist and Bayesian settings, both following the literature in this field (such as de Pooter, Ravazzolo and van Dijk (2007)) in addition to introducing some new combination approaches. The forecasts from individual models and combination schemes are evaluated in a number of ways; preliminary results show that model averaging generally assists in mitigating model risk, and that simple combination schemes tend to outperform their more complex counterparts. Such findings carry significant implications for central-banking analysis: a unified approach towards accounting for model uncertainty can lead to improved forecasts and, consequently, better decisions.