Forecasting the comovements of spot interest rates

Time-varying covariance models are compared in the French and German interest rate markets of the pre-euro period. A bivariate, asymmetric dynamic covariance model with level effect best characterizes the in-sample variance-covariance dynamics. Model comparison using economic loss functions raises some doubts with alternative models performing similarly. Out-of-sample results show that the variance-covariance matrix is best forecasted using a VECH model with level effect but no volatility spillover, not entirely confirming the in-sample evidence. Simple models using exponentially-weighted moving averages of past observations perform similarly to the best bivariate model. Thus, some features required to obtain a good in-sample fit do not have additional out-of-sample forecasting power due to overfitting.