Temporal Fusion Transformer for Multi Horizon Bitcoin Price Forecasting

Abstract

Bitcoin is one of the cryptocurrencies that have a large number of holders and a high transaction volume. Its price has experienced major fluctuation since its inception, resulting in relatively high price volatility. This has been a challenge for researchers over the past decade in forecasting its future price. In this study, we aim to address this problem by utilizing a neural network model called the Temporal Fusion Transformer (TFT). The TFT model uses the concept of multi-horizon forecasting to learn from historical data and predict future prices based on multiple time steps. Due to the numerous factors influencing Bitcoin's price, our study incorporates additional predictive data, such as Twitter sentiment, trends, and seasonality. The proposed method is trained using the TFT model with historical Bitcoin data from 2014 until 2020, then we performed transfer learning by using different past covariates as inputs. As a result, the proposed method in this study demonstrated the best performance, with a MAE of 0.05, a RMSE of 0.07, a MAPE of 5.92%, and a quantile loss of 0.03, surpassing other approaches. This makes it the state-of-the-art solution for highly volatile forecasting using a neural network model.