Cryptocurrency Price Forecasting Using Long Short-Term Memory (LSTM) with Twitter and Historical Data by Earl Timothy D. Malaki and Demelo M. Lao
Paper presented during
National Conference on Information Technology Education (NCITE2019) organized by the Philippine Society of Information Technology Educators Foundation, Inc. (PSITE) on 17-19 October 2019 held at Father Saturnino Urios University in Butuan City, Agusan del Norte
Abstract
The rise of cryptocurrency trading caused an increase in Twitter activity from field experts. Activities include sharing market predictions, opinions, and sentiments. As past studies have shown, Twitter data can be used to predict the market movement of financial instruments. Using these knowledge, we created a predictive model that predicts Bitcoin’s t+1 closing price using Bitcoin twitter and historical data. Tweets and historical data worth 181 days were gathered. Tweets were assigned sentiment score using vaderSentiment. The dataset included the following features in daily format: sentiment score, number of retweets and favorites, opening/highest/lowest/closing prices, trading volume, and market cap. A long short-term memory (LSTM) recurrent neural network was iteratively ran with different parameters and train-test methods. The best performing model achieved an RMSE and R2 of $110.57 and 0.98, respectively. It utilized 7 out of 9 features and used a walk-forward train-test method. Significant performance improvement was observed when using selected features, and even more so when using walk-forward method. Results have shown that it is possible to forecast Bitcoin’s t+1 closing price using twitter and historical data. The experimentation also has shown the importance of feature selection, and the benefit of walk-forward over a standard 70-30 train-test split method.