Should economic theories guide the machine learning model in forecasting exchange rate?

This study investigates whether integrating economic theory into the machine learning model by imposing monotonic constraints can improve the predictability of exchange rates. The black-box machine learning models have been praised for their predictive power in the empirical literature, leaving the question of the usefulness of economic theory unanswered. Using the tree-based model, we can impose the monotonic constraints implied by the economic theories on the possibly nonlinear relationship between the exchange rates and predictors. The empirical analyses suggest that the constrained models (with theory) often outperform those without constraints (without theory) in terms of statistical accuracy. In an experiment to examine the economic value, the currency portfolios based on these model predictions also deliver better risk-adjusted returns than the commonly used strategies, such as carry trade and momentum. The findings suggest that economic theories should be combined into the tree-based machine learning model for more accurate exchange rate forecasts.