Comparison of Activation Functions in Recurrent Neural Network for Litecoin Cryptocurrency Price Prediction

Abstract

The rapid advancement of information technology and digitalization has significantly transformed the financial sector, particularly with the emergence of cryptocurrencies characterized by high price volatility and complex movement patterns. Accurate price prediction of these crypto assets is essential to support investment decision-making and risk management. This study aims to compare the performance of six activation functions ReLU, Tanh, Sigmoid, Softplus, Swish, and Mish in a Simple Recurrent Neural Network (RNN) model for predicting the price of Litecoin, a widely traded cryptocurrency. Using historical daily closing price data from May 2020 to April 2025, the data were preprocessed through Min-Max normalization and sliding window sequence formation to fit the RNN input requirements. Each activation function was applied in the RNN model under consistent training conditions, and model performance was evaluated using Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), and coefficient of determination (R²). Results indicate that the Swish activation function outperforms others by achieving the lowest RMSE of 4.58 and the highest R² score of 0.9578, demonstrating superior prediction accuracy and stable convergence. Tanh also showed competitive results, while Sigmoid and Softplus performed less effectively. In conclusion, Swish is recommended as the most suitable activation function for RNN-based cryptocurrency price forecasting due to its balance of accuracy and computational efficiency.