Dynamic model averaging for daily forex prediction: A comparative study

Abstract

There are primary concerns in forecasting of time series especially in financial and economical data; model uncertainty in selecting predictors, manipulating the interested parameters in time-varying state vector, and to achieve the improvement on predictive performance in terms of Mean Square Error (MSE) and Mean Absolute Deviation (MAD). Theoretically, the algorithm in dynamic model averaging (DMA) is considered to be handling all the concerns simultaneously. In this work, we apply DMA and other models that are transformed from general DMA to predict daily percentage change in log-return of three most traded forex, i.e., JPY-USD, EUR-USD, and GBP-USD. According to empirical results, structural break model like AR(4) and TVP-AR(4) perform the best in forecasting USD-JPY. Moreover, in order to get the best performance to forecast EUR-USD, it is better to use parsimonious model. In other words, models which allow coefficients to evolve via random walk is preferable. Finally, we found that predictive performance of much complicated methods in DMA and DMS is extremely consistent over time in both forecasting exercises h = 1 and h = 4.