Time-Series Transformer for Predicting Bitcoin Price

Abstract

Bitcoin's high volatility poses significant challenges for short-term price prediction, making it a critical area of study for financial forecasting. Traditional models such as Long Short-Term Memory (LSTM) networks often encounter difficulties in handling long-range dependencies and non-stationary data, limiting their predictive accuracy under volatile conditions. This study introduces the Time-Series Transformer (TST) as a novel approach to predict Bitcoin's short-term prices. By leveraging self-attention mechanisms, TST effectively captures complex temporal patterns in historical Bitcoin data, including prices and trading volume. The data was segmented into fixed-length windows to facilitate model training and testing. Evaluation metrics such as Mean Squared Error (MSE), Mean Absolute Scaled Error (MASE), and R-squared (R²) demonstrated TST’s superior performance over LSTM, particularly during periods of high market fluctuation. Furthermore, TST exhibited notable computational efficiency when working with large datasets, underscoring its scalability. These findings not only highlight TST’s potential for enhancing cryptocurrency price prediction but also pave the way for future research integrating external data sources and exploring further model enhancements for more robust financial forecasting.