Deploying a Multi-Model Forecasting System for Bitcoin Prices: Bridging Statistical Forecasting and Deep Learning Innovations

Abstract

In this study, we investigate and compare several forecasting models for predicting Bitcoin market prices using historical data sourced from Nasdaq Data Link (formerly Quandl) spanning from 2016 to 2025. Our analysis evaluates traditional time series methods—such as ARIMA and Holt Winters exponential smoothing—alongside modern machine learning and deep learning techniques including LSTM, Prophet, XGBoost, SVR, Random Forest, and GRU. Performance was assessed via metrics such as RMSE, MAE, MAPE, sMAPE, directional accuracy, and R-squared. Our experiments reveal that while classical methods (e.g., ARIMA and Holt Winters) exhibit large estimation errors and limited explanatory capacity, advanced neural network architectures— particularly the GRU—demonstrate superior accuracy with an RMSE of 2,505.84, MAE of 1,760.93, MAPE of 2.79%, and an R-squared of 0.99. The best-performing model (GRU) was deployed as a web application on PythonAnywhere, providing real-time forecasts through an interactive dashboard. This deployment not only validates the predictive efficacy of the GRU model but also offers a practical tool for investors and financial analysts to monitor and predict Bitcoin price movements using reliable Nasdaq data.