Modelos de predicción estocástica para bitcoin : una evaluación de métodos y desempeño

Abstract

This research focuses on forecasting Bitcoin (BTC) prices using statistical models, including LSTM, GRU, SVR, decision trees, Random Forest, and XGBoost. We evaluate their performance in terms of R2, RSME, MAPE, Lin Concordance Coefficient (CCC), and Explained Variance Score—metrics selected for their ability to assess regression models. We utilized BTC closing price data from 2014 to 2023, subjected to preprocessing involving cleaning, optimization, and data engineering. The models, initially unoptimized, were enhanced through hyperparameter tuning and specialized statistical techniques such as cross-validation, L1-L2 regularization, Bayesian and genetic optimization. The results highlight XGBoost as the optimal model with the incorporation of iterative hyperparameter tuning, Bayesian optimization, and nested cross-validation. It achieved outstanding values in all evaluated metrics: RSME of USD 30.45, MAPE of 0.09%, R-squared of 1.0, Lin Concordance Coefficient, and Explained Variance Score of 1.0 in each case.