Evaluation of Price Prediction Models for Cryptocurrencies based on convolutional neural networks trained on Candlestick Charts

Abstract

In the past few years, there has been a growing interest in cryptocurrencies. However, the risk of incurring losses is high due to their large price fluctuations. Therefore, we want to reduce this risk by predicting the rise and fall of their prices. In this study, we use a convolutional neural network model trained on candlestick charts to make price predictions. In this experiment, the system was trained on the image pattern data of a set of five candlesticks, and predictions were made on whether the price would go up or down. The novelty of this research is that we apply the stock price prediction method using visual candlestick patterns, which has been empirically judged, to virtual currency prediction based on their visual pattern transition model with deep learning. The model trained on the data from 1-minute intervals gave the best results, with a predictive accuracy of 58.69% and a bankruptcy probability of only 1.678%.