Cryptocurrency Price Prediction With Multi-task Multi-step Sequence-to-Sequence Modeling

Abstract

In the scientific and commercial worlds, predicting cryptocurrency prices over time has gotten a lot of interest. Most related studies use variations of Recurrent Neural Networks to forecast the next value of a single coin due to the temporal nature of the challenge. As a result, determining how effectively such a model would perform across various tasks (cryptocurrencies), several future timesteps, and forecasting horizons will be difficult. This paper proposes a multi-task and multi-step sequence-to-sequence model that is trained jointly on 22 cryptocurrencies' time series. Our findings show the value of sequence-to-sequence modeling for future predictions, as well as the significant improvements in accuracy and training time that can be achieved by using a single multi-task model rather than numerous distinct models for each task.