Bitcoin candlestick price prediction with recurrent neural network

Abstract

Bitcoin is a high-risk asset with a potentially high return. Predicting Bitcoin candlestick, i.e., open, high, low, and close (OHLC) prices, can help investors make trading decisions. The objective of this study is to develop a neural network model to predict the candlestick prices of Bitcoin for the next period. Additionally, this study investigates methods to enhance the model's forecasting performance by feature transformations, specifically data normalization. This study employs two neural network algorithms, Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU), to forecast daily Bitcoin OHLC prices. To enhance the model's performance, we compare sliding window normalization with whole set normalization techniques. The normalization techniques investigated for both whole set and sliding window data include z-score normalization, min-max normalization, and relative change normalization. Furthermore, this study compares two candlestick prediction methods, namely using OHLC prices and using candle wick (CULR) to predict OHLC prices. The models use historical OHLC prices over several days to predict the next day's OHLC prices. The results indicate that the best-performing model is the OHLC method using GRU algorithm with sliding window z-score normalization, which achieves an MAPE of 1.95% and an RMSE of 767.71. Moreover, the sliding window normalization generally outperforms the whole set normalization for both LSTM and GRU models in terms of RMSE and MAPE. Regarding the candlestick prediction methods, there was no significant difference in their performance in terms of accuracy and forecasting error. However, our results suggest that the OHLC method performs slightly better than the CULR method.