Cryptocurrency price prediction using machine learning technologies: LSTM vs KAN

Machine learning (ML) technologies predict the complex nonlinear behavior of the volatile cryptocurrency market with high accuracy. However, there is an urgent need to develop models that not only provide high forecasting accuracy, but also demonstrate reliable interpretability of the obtained results. As one of the most advanced artificial intelligence (AI) technologies, the Kolmogorov-Arnold network (KAN) acts as a promising alternative to traditional ML technologies, opening up new opportunities for improving the interpretability of AI models. The purpose of the study is to conduct a comparative analysis of the accuracy of cryptocurrency price forecasting models Bitcoin (BTC), Ethereum (ETH), Litecoin (LTC): 1) based on a recurrent neural network (LSTM); 2) based on the Kolmogorov-Arnold neural network (KAN), taking into account various adequacy indicators (accuracy metrics) of the models. The results show that for BTC and ETH cryptocurrencies, the KAN forecasting model outperforms the LSTM forecasting model in all considered adequacy indicators, which indicates a high forecasting potential of the KAN model. From a practical point of view, the obtained results are of undoubted interest in the development of an effective investment policy in the cryptocurrency market based on accurate cryptocurrency price forecasts. Accurate cryptocurrency price forecasting is important for market participants for a number of reasons, such as building trading strategies, risk management, price discovery, market sentiment analysis, and business applications. Understanding the dynamics of major cryptocurrencies is also crucial for policymakers, given their impact on investment strategies, market efficiency, and regulatory oversight.