Investigating the Role of Activation Functions in Predicting the Price of Cryptocurrencies during Critical Economic Periods

Laszlo Vancsura – Tibor Tatay – Tibor Bareith

Virtual Economics, Vol. 7. No. 4. (2024)

Abstract

Accurate cryptocurrency price forecasting is crucial due to the significant financial implications of prediction errors. The volatile and non-linear nature of cryptocurrencies challenges traditional statistical methods, revealing a gap in effective predictive modelling. This study addresses this gap by examining the impact of activation functions on neural network models during critical economic periods, specifically aiming to determine how optimising activation functions enhances accuracy in neural network models, including RNN, GRU, LSTM, and hybrid architectures. Using data from January 2016 to June 2022—encompassing stable periods, the COVID-19 pandemic, and the onset of the 2022 Ukraine conflict—we analysed price trends under various market conditions. Our methodology involved testing three activation functions (ReLU, sigmoid, and Tanh) across these models. Both univariate and multivariate analyses were conducted, with the latter incorporating additional metrics such as opening, highest, and lowest prices. The results indicate that optimising activation functions enhances prediction accuracy. Among the models, GRU demonstrated the highest accuracy, whereas RNN was the least efficient. Multivariate models outperformed univariate ones, highlighting the benefits of incorporating comprehensive data. Notably, the Tanh activation function led to the greatest improvements, particularly in underperforming models such as RNN. These findings underscore the critical role of activation function selection in enhancing the predictive power of neural networks for cryptocurrency markets. Optimising activation functions can lead to more reliable forecasts, facilitating better trading decisions and risk management. This study highlights activation functions as key parameters in neural network modelling, encouraging further exploration. Future research could investigate different economic periods and cryptocurrency behaviours to assess model robustness. Additionally, examining a broader range of cryptocurrencies may reveal whether the benefits of activation function optimisation are consistent across various assets. Incorporating external factors such as macroeconomic indicators or social media sentiment could further enhance models and improve forecasting accuracy.

Keywords:

cryptocurrencies; deep learning; forecasting; activation functions; COVID-19; Russian-Ukrainian conflict

https://doi.org/10.34021/ve.2024.07.04(4)