| dc.description.abstract | This study revisits the conclusions of Meese and Rogoff (1983), which suggested that the random walk model is the best for short-term exchange rate forecasting. Using monthly data and various machine learning models—including shrinkage models, principal component models, random forests, and deep neural networks (DNN)—we forecast the US dollar exchange rates against major currencies (yen, euro, Australian dollar, Canadian dollar, and British pound) and compare them with the random walk model.
Empirical results show that machine learning models outperform the random walk model for one-month-ahead forecasts. However, their predictive power decreases for two and three-month-ahead forecasts, partially supporting Meese and Rogoff′s conclusions. LASSO analysis reveals that most important variables are global interactions, highlighting the significance of global factors in exchange rate predictions. The Clark-West test also confirms that most machine learning models are significantly better than the random walk model for short-term forecasts, though their advantage diminishes over longer horizons.
In summary, machine learning models show significant potential for short-term exchange rate forecasting, particularly for one-month-ahead predictions. However, their accuracy declines over longer periods, not consistently outperforming the random walk model. Future research should explore additional variables and improved model structures to enhance long-term forecasting accuracy. | en_US |