CRYPTOCURRENCY PRICE PREDICTION USING TRANSFORMER: A DEEP LEARNING ARCHITECTURE

Abstract

In recent years, price prediction for cryptocurrency has gained so much attention as there are a lot of investors, spectators and consumers in the market and are interested to know where to spend their money for profitable trades. Cryptocurrency is attracting customer in reshaping the finance structure because of its fame and acceptance. Bitcoin is the most popular and transparent cryptocurrency over the internet. Due to the rapid fluctuation in bitcoin prices, uncertainty, dynamic features, etc. there comes a need to predict its price as it varies rapidly in short interval of time. Many deep learning model has been implemented for price prediction but attention mechanism is proved to be more efficient and has sought focus in recent years. The aim of our work is to develop a trained and efficient machine model which can predict bitcoin price if we input huge amount of data with a good computational time and power. In our study, we proposed transformer architecture which implements attention layer. Our dataset contains features which are related to bitcoin price and has data of over ten years recorded daily. Features in the dataset are dependent on each other which can affect the price level for the upcoming time. The architecture is used with two types of deep learning models which are LSTM and GRU. Both these model are used as LSTM-with-attention and GRU-with-attention. Finally, we compared both these models on the basis of four error metrics which are mean absolute error (MAE), mean square error (RMSE), maximum error (ME), and root mean squared error (RMSE). After comparing results, it has been observed that the transformer model GRU-with-attention gives less error rate i.e. better estimation than LSTM-with attention. The results obtained might have inferences on implementation of more complex time series problems with deep neural networks.