A neural network-driven dynamical system model for cryptocurrency price prediction

Abstract

This paper introduces an enhanced method for forecasting cryptocurrency prices by employing a dynamical system model powered by Neural Networks. While conventional Neural Networks are commonly utilized for time series prediction, they often fall short in handling the highly volatile nature of cryptocurrency data, which is prone to sudden fluctuations. To address these challenges, this study leverages a different methodology known as the 2DSL approach, which decomposes cryptocurrencies' price signals to better capture underlying trends. Although the original 2DSL model has improved performance over traditional methods (e.g., LSTM and GRU), it still encounters instability in its predictions, mainly when dealing with high-frequency fluctuations in the data. To tackle this problem, we introduce an enhancement by merging high-frequency decomposed signals, stabilizing the model, reducing the number of input data points, simplifying the Neural Network, and decreasing computational complexity. This improved approach allows the model to focus on more stable, low-frequency components that better reflect long-term market trends. To evaluate the effectiveness of the enhanced 2DSL model, historical Bitcoin data was used as a test case. The numerical results demonstrate that the model provides more accurate and stable price predictions than conventional Neural Networks, with significant improvements in short-term and long-term forecasts.