Extracting cryptocurrency trading signals from raw financial data using technical indicators and deep learning

The purpose of this study is to investigate deep learning methods for predicting cryptocurrency, specifically Bitcoin, trading decisions considering different oversampling and augmentation methods, and using different feature groups as input, extracted using technical indicators. For this purpose, financial time series data is labelled as three classes namely, buy, sell, hold, by a window sliding approach. However, this labeling approach gives rise to imbalanced data set which is problematic in deep learning setting. In order to address imbalanced data set problem, three oversampling methods which are Direct Copying, SMOTE, ADASYN, and two augmentation methods, jittering and time warping, are used for increasing the size of data sets that belong to minority classes. Moreover, technical indicators are included in this study to extract financial data features to be fed into deep learning models. Feature selection and ablation study are performed on the extracted features. Six different feature groups are created which can be listed as; all indicators, top 15 indicators selected using mutual information metric, and momentum, volume, volatility and trend indicators. The proposed deep learning architectures, convolutional neural network and attention mechanisms with two different encoder architectures, are trained with different combinations of oversampling and augmentation methods, and feature groups. Results are promising where the best macro average F1 score of 56.48 is achieved by the model generated using attention deep learning architecture which is fed with top 15 indicators where SMOTE oversampling technique is applied.