MLP-based Learnable Window Size for Bitcoin price prediction

Abstract

Over the past few years, Bitcoin price prediction has been changed to a big challenge for investors on cryptocurrencies. In this regard, Neural Networks as a strong structure for regression analysis would play an important role to make a precise prediction. While several leading researches in this field considered the features affecting the price of bitcoin by a fixed number of past days, a new method entitled Learnable Window Size (LWS) is presented for smartening the number of days intended to predict the price of Bitcoin the next day. This paper implements a primary deep neural network, based on the observed Bitcoin price trend in the past days and its fluctuations, to predict the best window size. Then, the secondary deep neural network predicts the price of Bitcoin according to the predicted window of the first step. The dataset of this paper is included Google, Blockchain, and Bitcoin market data. Evaluations have shown that based on the Prediction Hardship Factor (PHF), a new criterion which has been proposed to describe the degree of difficulty of prediction, this method has been able to get the minimum error under a normal situation which is superior in comparison to the well-known methods such as Support Vector Regression and ARIMA.