Classifying short-term forex trading opportunities with machine Learning

In this study, it was desired to investigate if short term movements in the forex market can be classified with success. Using hourly price data from 2010-2024 (with 2024 as the validation set), data labels were constructed for each hour, and technical indicators along with chart patterns and custom variables were constructed and used as independent variables for the classification problem. Time series clustering using k-means and dynamic time warping was also utilized as a feature to characterize the shape of the data to understand its future short-term price movement behavior. Several classifiers were tested, with XGBoost yielding the best results. It obtained a class-weighted accuracy of 82.6% and F1 score of .826. Using this classifier as a way to make simulated trading decisions, a system was constructed and generated 11% annualized returns or higher using the validation dataset, with a maximum portfolio drawdown of 10%.