Optimizing Bitcoin Price Prediction: Multivariate LSTM Triumphs

Abstract

This study aims to assess the validity and precision of employing a multivariate LSTM model compared to traditional models and stock analysis techniques for predicting the price of the cryptocurrency BTC. The research incorporates a feature elimination technique to optimize price predictions across various time intervals by removing non-essential and redundant features, including economic factors. In the case of BTC, with a finite total supply of 21 million coins, an increase in popularity generally leads to a surge in price. To gauge BTC’s popularity, tweet frequency and Google search trends were considered as input factors. Additionally, traditional indicators like USD, Gold and the Volatility Index (VIX) were used to measure the stock market atmosphere. The LSTM model’s performance was benchmarked against other models such as RNNs, ANN, SVR and ARIMA. The LSTM model exhibiting superior learning in multivariate data, achieving an RMSE score of 268.83.