Prediction of Bitcoin Price Based on Transformer, LightGBM and Random Forest

Abstract

Contemporarily, cryptocurrency have been established and widely recognized as an alternative method of exchange currency. As computer technology advances, the trading of cryptocurrency is increasingly viewed as a popular and lucrative form of investment. On the other hand, the volatile nature of cryptocurrency results in sudden and unpredictable price fluctuations. Consequently, the necessity for developing a precise and dependable predictive model for portfolio management and optimization is acknowledged. At present, RNN and other deep learning models are commonly utilized in predicting stock prices, ensemble learning methods including Random Forest and Xgboost have been widely implemented and researched in the realm of investments, engagement in trading activities in various fields including gold and other markets. However, research on cryptocurrency investment is still insufficient and the practical application of the mentioned models has not yielded satisfactory returns. As machine learning and artificial intelligence continue to advance, the enhancement of machine learning models' mechanisms could significantly boost the accuracy in predicting cryptocurrency prices. Hence, the utilization of Transformer architecture alongside various established models was opted for in the prediction of Bitcoin's price. The attention mechanism embedded in the Transformer model has proven to be effective in anticipating the value of assets undergoing significant price shifts. Meanwhile, the difference in the accuracy of short-term and long-term price prediction of each model may also help and guide cryptocurrency investors in model selection and development during investment and market analysis.