Cryptocurrency Price Distribution Prediction Using Neural Networks

Our final project, a paper which uses neural networks to provide Bitcoin and other cryptocurrency traders with good buy/sell signals. This is our final project for CSC413H1, Winter 2021, taught by Jimmy Ba and Bo Wang at the University of Toronto. The paper can be found under "Cryptocurrency Price Distribution Prediction Using Neural Networks.pdf".

Given only a one-month time frame for working on the project, we hope to continue contributing towards achieving higher accuracy and lower mean error. After attempting many different approaches and conducting numerous iterations throughout the project, we realise that there are more studies and algorithms to learn and implement.

If you are experienced in working with machine learning algorithms for trading, feel free to contact us for additional resources that may help grow this project!

Key take-aways

• complex models with a diverse set of layers perform better
• weight regularization is an effective technique for reducing overfitting in this task
• combining different types of technical indicators can improve performance
• our best model, combining CNN, RNN and MLP layers, made a 16% improvement in prediction accuracy over other research
• the random walk theorem is not necessarily applicable to crytocurrencies, as previously suggested by other papers

Abstract

Price prediction in the financial market is attractive for both holders and traders due to its volatility and uncertainty. Many have tried to analyze the future of the financial market by using methods and models from machine learning. In this paper, we construct several neural network models to predict sequential data for cryptocurrencies, specifically Convolutional Neural Network (CNN), Recurrent Neural Network (RNN), CNN-RNN Multilayer Perceptron (CNN-RNN-MLP), Long Short-Term Memory with Attention (LSTM-Attention), and InceptionNet. The results showed that CNN-RNN-MLP performs the best among other models.

Introduction

Financial time series prediction is a major topic within the finance industry. It is a difficult task due to the uncertainty of the financial market. In recent years, the strategies of predicting financial assets have expanded. Researchers have proposed many state-of-the-art methods for stock and cryptocurrency predictions, such as sequential prediction using the CNN and LSTM. However, current research has not proposed models which give traders buying and selling points. Current research also does not explore newer machine learning model types and does not take advantage of technical indicators used by professional human traders. In this paper, we are going to focus on the forecast of cryptocurrencies using several neural network techniques. First, we train our models with Bitcoin data, where the output of our models is a percentile of our choosing of the distribution of Bitcoin prices of the next 12 hours, thus providing signals for traders on when to buy and sell Bitcoin in the short term. Then we use the top performance models to perform extensive experiments on different cryptocurrencies and on different percentiles. Lastly, we compare our model architectures against those from other research.

Quick Guide

1. Download data using data_downloads/DownloadBinanceData.py.
2. Generate a data set using data_set/CreateDataSet.py.
3. If you want the model to predict whether a percentile will be higher or lower than the previous day's mean price, set BINARY_PREDICTION to True in util/Constants.py.
4. Train the models in the models folder.
5. Enjoy!

Machine Learning Models and Layers Used

• CNN
• LSTM
• CNN + LSTM + MLP
• LSTM + Attention
• InceptionNet

Technologies Used

TensorFlow 2, Keras, Python