A Machine Learning Modeling for Bitcoin Market Price Prediction based on the Long Short Term Memory Recurrent Neural Network

Abstract

Background/Objectives: Machine learning based on Neural Network has integrated usages in a variety of fields such as translation, finance, distribution, and medical world as well as cognition. This study shows Recurrent Neural Network Learning Model on the basis of LSTM, which analyzes the previous prices of a cryptocurrency, Bitcoin and predicts the next one.

Methods/Statistical analysis: This model indicates the actual and predicted prices of Bitcoin for 81 days in the way that it has learned the former prices for 30 days and then anticipates the next day price. Regularized data set for Modeling is divided into test data set and training data set at the rate of 1:9. The latter set is once again separated into training data and verification ones. Machine Learning of this study needs to use Neural Network library, Keras framework.

Findings: To fit the model is to look for the model’s weight by optimizing the process, while using the training data. In this paper, fit function’s batch size is 11 and epochs is 30. As learning gets processed more repetitively, the loss decreases more monotonously, and then it converges to more regular value. That is, it means there is no overfitting.

Improvements/Applications: As the result of the experiment, the machine learning proposes not only that after analyzing the graphs of error rates and weight change rates, weight converges towards a particular one, but also that as learning goes over, the processing efficiency of its neural network gets better. 

Keywords: Block Chain, bitcoin, price, Recurrent Neural Network, LSTM, prediction