Do the combination models perform better than the fundamental models in forecasting the exchange rates?

This study examines the predictability of the simple average combination model and the inverse average error combination model in forecasting the out-of-sample EUR/USD, GBP/USD, and JPY/USD exchange rates from 1st July 2019 to 30th June 2020. Out of the three currency pairs examined, both of the combination models only show evidence in forecasting the JPY/USD exchange rate under the 1-month horizon, in which the absolute values of their z-statistics are smaller than the two-tailed 5% significance level critical value, 1.96. In terms of the forecast performance comparison of the simple average combination model, the inverse average error combination model, the PPP model, the uncovered interest rate parity model, the real interest differential model, and the Taylor rule fundamental model, none of them consistently outperforms the others. Nonetheless, I find that the inverse average error combination model overall produces lower average absolute errors than the simple average combination model. There is also evidence showing that the inverse average error combination model generates smaller forecast deviations as compared to the PPP model, the uncovered interest rate parity model, the real interest differential model, and the Taylor rule fundamental model, respectively for different currency pairs under different forecast horizons.