We establish the out-of-sample predictability of monthly exchange rates via machine learning based on 70 predictors capturing country characteristics, global variables, and their interactions. To better guard against overfitting in our high-dimensional and noisy data environment, we move beyond “off-the-shelf” implementations of machine learning methods to induce adequate shrinkage. The resulting forecasts consistently outperform the random walk benchmark, which has been a focus of research and proven difficult to beat. Out-of-sample exchange rate predictability is linked to the state of the economy, with stronger predictability during business-cycle recessions, while model interpretation tools reveal that a number of country characteristics are important when they interact with global variables. Machine learning forecasts are also economically valuable to investors, as they substantially improve the performance of portfolios that invest in foreign currencies, especially since the Global Financial Crisis, when carry trade portfolios suffer sharp declines in profitability. From an asset pricing perspective, our machine learning forecasts paint a highly plausible picture of the conditional price of currency risk, with the price of risk markedly increasing during periods of financial and economic stress.

Keywords: Panel predictive regression, Elastic net, Deep neural network, Asset allocation, Conditional price of risk

Continue reading here.