Forecasting exchange rates using an optimal portfolio model with time varying weights

Abstract

This paper presents a mean variance based model of exchange rate determination and forecasting using the return differential of an optimal portfolio composed of money, bond, and stock market returns. We use the simple OLS estimation technique for the estimation and a recursive rolling regression technique to generate the out-of-sample forecasts. We employ an autoregressive technique to estimate the mean returns and time varying variance covariance matrices to generate time varying portfolio return weights. The out-of-sample forecast analysis, using the CW statistic suggests that our Optimized Uncovered Rate of Return Parity model outperforms the naïve random walk model in forecasting one month ahead nominal exchange rates for all the countries in the study. The results also show that the un-optimized model is also able to outperform the naïve random walk in all the countries at one month ahead forecasting horizon. These findings imply that the inclusion of the three market variables in modelling exchange rates improves the forecasting ability of exchange rate models.