Exploring Normalization Techniques in Neural Networks for Bitcoin Candlestick Price Prediction

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. This paper aims to explore various data normalization techniques in neural networks to enhance candlestick price prediction. In this study, the sliding window normalization techniques were compared with the whole set normalization techniques for forecasting the daily Bitcoin OHLC prices using two neural network algorithms: Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU). The investigated normalization techniques for both the whole set and the sliding window set included z-score normalization, min-max normalization, and relative change normalization. Historical OHLC prices over several days were used to predict the next day's OHLC prices. The results show that the sliding window normalization techniques outperformed the whole set normalization techniques in terms of RMSE and MAPE with the best technique being the GRU algorithm using the sliding window relative change normalization achieving MAPE of 2.25% and RMSE of 870.52.