Measuring the Performance of Machine Learning Forecasting Models to Support Bitcoin Investment Decisions.

This research proposed machine learning forecasting models to support bitcoin investment decisions based on bitcoin price and trade volume from 2019 to 2021. The moving average crossovers of 5, 30, and 90 daily closing prices and their variances were inputs loaded into decision tree, random forest, and extreme gradient boosting (XGBoost) techniques to forecast bitcoin investment strategies, including market trends, actions, and holding amounts. The research also measured the models' performance based on accuracy, precision, recall, F1-score, and area under the curve-receiver operating characteristics (AUC-ROC). The results indicated that the XGBoost is the most efficient model: (1) trend (0.930 accuracy, 0.930 precision, 0.930 recall, 0.929 F1-score, and 0.983 AUC-ROC); (2) action (0.985 accuracy, 0.985 precision, 0.985 recall, 0.985 F1-score, and 0.998 AUC-ROC); and (3) amount (0.987 accuracy, 0.987 precision, 0.987 recall, 0.987 F1-score, and 0.997 AUC-ROC). The random forest achieved the second most efficient model, while the decision tree provided the lowest forecasting results. Since the bitcoin investment market in 2022 is significantly different from the previous two years due to several negative factors, the research further validated the models' performance with an unseen data set comprising 275 days of bitcoin market prices from January 1 to October 2, 2022. All the models suggested that investors hold with half the investment consistent with the investment market in 2022. Furthermore, although the decision tree and XGBoost models forecasted the investment trend for most days as up, the random forest forecasted the trend as sideway, consistent with the 2022 trend. [ABSTRACT FROM AUTHOR]