Machine Learning Algorithms in Predicting Prices in Volatile Cryptocurrency Markets

Abstract

This study aims to develop a predictive model for cryptocurrency prices in highly volatile markets. The methodology includes an exploratory data analysis, followed by designing and implementing machine learning (ML) algorithms, focusing on the Long Short-Term Memory (LSTM) neural network. The model's performance was optimized through hyperparameter tuning, and its stability was validated using an analysis of variance (ANOVA). We conducted a benchmark comparison with other ML approaches. Our LSTM model achieved an R² of 99.41% on the first day of prediction and maintained an accuracy above 97% up to the seventh day, demonstrating its robustness even for extended forecasts. During training, the LSTM model reached an RMSE of $1,187.14 and a MAPE of 2.20%, with the MAPE consistently remaining below 10% during the validation phase. For seven-day forecasts, the model recorded an RMSE of $5,038.46 and a MAPE of 6.83%. In comparison, alternative models such as Support Vector Machines (SVM), Extreme Gradient Boosting (XGBoost), and Random Forests exhibited significantly higher error rates; for instance, XGBoost recorded an RMSE of $17,849.66 and a MAPE of 27.74%. Overall, these findings highlight the superior performance of the LSTM model in addressing the challenges of cryptocurrency price forecasting.