Forecasting logarithmic returns on bitcoin based on the Informer model

Abstract

Recently, quantitative trading techniques applied in financial research have become increasingly sought after. Quantitative trading refers to the use of statistics and computer techniques to aid trading decisions. Bitcoin has attracted a large number of investors to invest in it due to its decentralised nature, anonymity, and total number of 21 million pieces. This paper wishes to profit from investing in Bitcoin. This paper predicts the logarithmic return of Bitcoin based on the Informer model. Because of the high volatility of Bitcoin, this paper shortens the prediction period of the Informer model from the traditional 24 days to 1 day. In addition, this paper introduces the technique of migration learning, where models trained on five tech company datasets are migrated to Bitcoin's dataset for training tests. This compensates for the small Bitcoin dataset to some extent. In this paper, MSE, MAE, and R-squared were used as the evaluation metrics with MSE of 0.5678, MAE of 0.5087, and R-squared of 0.232313938. The results show that the Informer model's short-term forecasting ability is validated. The value of this paper is to provide Bitcoin investors with a possible method to aid trading decisions.