Study on Improving the Prediction Power of Short-term High-volatility Cryptocurrency Price Using Deep Learning with Regime Transformation Labeling of Time Series and Image Transformation

We propose a framework to improve the prediction power of highly volatile and relatively short time series data, which works by transforming the target value to target regimes evaluated from time series information. Regime transformation labeling of time series in an window values employes a homogeneous regime switching model. Since regime labels exist for windows, we use computer vision deep learning models through image transformation with the Gramian Angular Field and data augmentation to improve prediction power. VGGNet, GoogleNet, and ResNet-50 are used for the deep learning models. For empirical analysis, we use price data from the cryptocurrency market, which is known for its high volatility and limited data availability. In particular, we assess performance using Trump-coin, Melania-coin and ChillGuy representative memetic coins. The experimental results show high prediction performance, as measured by ROC-AUC. Especially, ResNet-50 achieves ROC-AUC of 0.96 - 0.997. These results suggest that the framework proposed in this study can be effectively utilized in non-traditional asset classes with low initial trading volume or rapid market changes.