Abstract

Cryptocurrency price forecasting is inherently challenging due to pronounced volatility, non-stationary behaviour, and strong sensitivity to market sentiment. In this work, a comparative analysis is conducted between the recurrent Long Short-Term Memory (LSTM) network and the attention-based Time Series Transformer (TST) for predicting Bitcoin prices. Daily Bitcoin price data are modelled using three feature configurations: (i) raw OHLCV variables, (ii) OHLCV combined with technical indicators, and (iii) an extended feature set incorporating public interest measured through Google Trends for the term Bitcoin. To examine temporal sensitivity, the TST model is evaluated across multiple lookback windows of 30, 60, and 90 days. Experimental results reveal that LSTM with OHLCV features achieves the lowest absolute forecasting error (MAE: 3,958; RMSE: 5,621), while feature enrichment degrades performance in both models. The TST demonstrates greater robustness across lookback window configurations, with its MAE varying within a narrower range (10,987–12,406) compared to LSTM under feature enrichment. These findings contribute nuanced insights into the conditions under which attention-based architectures offer practical advantages over recurrent models for cryptocurrency price prediction.

Authors

N. Valliammal, P. Hemashree
Avinashilingam Institute for Home Science and Higher Education for Women, India

Keywords

Bitcoin Price Forecasting, Time Series Transformer, Deep Learning, Google Trends, Cryptocurrency