| With abundant accessible data, applying machine learning models for cryptocurrency is a major area of interest to many researchers and investors alike. However, most supervised machine learning-based strategies fail to deliver profit. The labeling method for the training data remains one of the factors contributing to this problem. This study develops a new labeling method, long reward labeling, by determining trend reversal points and calculating the reward of taking a long position in relation to those trend reversal points. The results suggest that long reward labeling can deliver a profitable trading strategy for different time horizons and time frames. The baseline recurrent neural network (RNN) model yields predictions that match the true long reward values and are responsive to the spikes in the true values. |
