Traders are still dealing with the puzzle of currency prediction. For decades there has been a simple challenge and research conclusion. Most models cannot beat the random walk at predicting currency movements. The machine learning explosion allows traders to explore different methods for looking at the same data. In a new paper, "Machine Learning in Foreign Exchange", the author looks a number of prediction techniques including neural networks and tree-based models. It finds that a neural network approach outperforms linear and tree-based models and can do a good job of predicting cross-sectional excess returns. It can beat the random walk but the predictive power declines at longer-term horizons.
A non-linear approach does much better than simple linear models; however, many avoid using ML techniques given their more difficult explainability. The author uses local interpretability techniques like DeepLIFT and Layer-wise Relevance Propagation (LRP) as well as Shapley values for global interpretability. What should not be surprising is that the key factors we always know as important for explaining currency returns still hold. Momentum and carry are still the most important factors for making currency predictions, yet it is how and when they interact with currencies that matter. The non-linear influences dominate currency production. Use the same features but link them together in new ways.
Now, when looking at the set of models. It is not always clear why one approach does better than another. There clearly may be overfitting with some models. The feature importance also are complex. There are many carry features that can provide forecast value which seems odd. Nevertheless, there is a lot of good work her which needs further testing to further support currency predictions.
No comments:
Post a Comment