An Experiment on Predicting Cryptocurrency prices with LSTM Networks

This repository does not constitute financial advice and should not be taken as such. While cryptocurrency investments will definitely go up in value forever, they may also go down.

Problem Description

In this experiment I build an LSTM network in order to predict the price of Bitcoin based on scenario described at [1] and [3].

Data

The data about Bitcoin was extracted from coinmarketcap.com. The training data consists of multiple multivariate time series with "day" as the time unit that start from 2013-12-27 to 2017-09-24. The testing data has the same data schema as the training data and start from 2017-09-25 to 2017-11-27. Finally, the model was blind tested using data from 2017-11-29 to 2017-12-06.

LSTM Configuration

The LSTM model will use previous data to predict the next day's closing price of bitcoin. So we set how many previous days it will have access to equals to 20.

LSTM params:

• epochs = 100
• batch_size = 32
• num_of_neurons_lv1 = 50
• num_of_neurons_lv2 = 25
• activ_func="linear",
• dropout=0.5
• loss="mean_squared_error"
• optimizer="adam"

Results:

Mean Squared Error	Mean Absolute Error	Coefficient of Determination (R^2)
0.003	0.048	0.603

Blind Test:

Single day prediction from 2017-11-28 to 2017-12-05: 9669.40, 10035.91, 9830.83, 9993.55, 10367.92, 10770.16, 10858.19, 11521.98

The truth from 2017-11-29 to 2017-12-06: 9888.61, 10233.60, 10975.60, 11074.60, 11323.20 , 11657.20, 11916.70, 14291.50

Pearson correlation between single day prediction and the truth: 0.92

References:

• [1] Predicting the price of Bitcoin using Machine Learning here.
• [2] Automated Bitcoin Trading via Machine Learning Algorithms here
• [3] Predicting cryptocurrency prices with deep learning here
• [4] Multidimensional LSTM Networks to Predict Bitcoin Price here
• [5] Kaggle Dataset here