Machine Learning and Qualitative Variables in Bitcoin Price Prediction: An Empirical Evaluation of Model Performance

This study employs the problem of Bitcoin price prediction to evaluate the performance of forecasting methods. Traditional linear regression is compared with machine learning techniques, specifically neural net fitting and neural net time series, to assess their predictive accuracy. Given Bitcoin’s high volatility and susceptibility to social, political, and behavioral influences, the study also examines the impact of qualitative factors on prediction performance. In addition to quantitative variables, qualitative variables—such as the Fear and Greed Index and sentiment analysis metrics—are incorporated into the models to enhance forecasting robustness. The results indicate that traditional multiple linear regression yields relatively weak predictive performance, whereas neural net fitting demonstrates superior accuracy. Furthermore, models relying solely on quantitative variables underperform across all tested methods. In contrast, the inclusion of qualitative variables significantly improves prediction outcomes in all approaches. The study concludes that integrating qualitative variables enhances Bitcoin price forecasting accuracy, with neural net fitting emerging as the most effective method due to its lower mean squared error (MSE) values.