Decoding Bitcoin: leveraging macro- and micro-factors in time series analysis for price prediction

Area of Science:

• * Computational Finance
• * Data Science
• * Econometrics

Background:

• * Bitcoin price prediction is vital due to its market influence and volatility.
• * Previous research often used limited data or features.
• * A need exists for comprehensive models incorporating diverse data for accurate Bitcoin price forecasting.

Purpose of the Study:

• * To predict Bitcoin prices using time series analysis with an extensive dataset (>2,000 days).
• * To evaluate machine learning and deep learning frameworks for time series prediction.
• * To identify optimal window sizes and enhance prediction accuracy through diverse input features.

Main Methods:

• * Utilized time series analysis on a dataset spanning over 2,000 days.
• * Incorporated diverse input features: technical indicators, sentiment analysis (social media, news, Google Trends), macroeconomic indicators, on-chain data, and traditional financial assets.
• * Evaluated machine learning and deep learning models, focusing on bidirectional long short-term memory (Bi-LSTM).

Main Results:

• * The bidirectional long short-term memory (Bi-LSTM) model demonstrated significant predictive performance.
• * With a window size of 3, Bi-LSTM achieved a Root Mean Squared Error (RMSE) of 0.01824, Mean Absolute Error (MAE) of 0.01213, Mean Absolute Percentage Error (MAPE) of 2.97%, and R-squared of 0.98791.
• * Gradient importance and ablation tests confirmed the validity and influence of various input features.

Conclusions:

• * The proposed methodology effectively predicts Bitcoin prices by integrating a wide array of data categories.
• * The Bi-LSTM model proves robust, even when including outlier events like the COVID-19 pandemic.
• * Findings support informed investment decisions and policy-making in the cryptocurrency space.