Forecasting bitcoin price considering macro economic factors and media influence using bidirectional LSTM and random forest regressor as ensemble model

Abstract

The decentralized cryptocurrency has created many opportunities for secure and safe financial transactions with a bright prospect. The cryptocurrency market rapidly expands, leading to erratic price movements due to geopolitical, social, and other macroeconomic factors. As a result, the price of such cryptocurrencies changes every day. For our research, we limit our scope to predicting and forecasting bitcoin prices accurately. For predicting the trend of Bitcoin price, we considered two major fac tors: the consideration of various macroeconomic markets and the sentiment analysis of social media. Our contribution to this research was the volume of data that we collected for sentiment analysis for tweets which is approximately 85 millions. In addition, we considered the impact of the markets of AMD and NVIDIA which are the main tech companies that provide consumer level GPU that has a huge impact in cryptocurrency mining, which has never been considered before for predicting cryptocurrency prices and to improve our accuracy we used ensemble Random For est Regression with Bidirectional LSTM. In this case, we considered Twitter. We have used the Vader Sentiment Analysis model to calculate the sentiment scores (positive, negative, neutral, and compound). We have used four parallel Bayesian Optimized Bi-LSTM models, each with its input features, to combine their predic tions and train an ensemble Random Forest Regressor with those predictions. Then, we used the trained RFR model to pick the best forecast out of those four parallel Bi-LSTM models. Furthermore, we got the following results: MSE = 0.0021607, MAE = 0.0318709, R2 = 0.99909, and MAPE = 0.0038217. The findings were that Bidirectional LSTM functions better in prediction when we consider sentiment anal ysis and other macroeconomic factors(AMD, NVIDIA, S&P 500, NASDAQ, GOLD stock prices). Moreover, using RFR as an ensemble model, the accuracy is boosted significantly.