A deep learning-based approach combining technical indicators and market sentiment for multi-timeframe cryptocurrency investment decision support
Abstract
To help navigate the volatility of cryptocurrencies, this study proposed a leakage‑free deep learning approach utilizing over 50 technical indicators as well as the Crypto Fear & Greed Index (FGI) to forecast movement for five major cryptocurrencies; BTC, ETH, BNB, XRP, and SOL. For evaluation, the research used daily time series data from January 2018–April 2025 and tested six different neural network approaches and several traditional machine learning baselines using TimeSeriesSplit for a strict cross‑validation methodology. In addition, class imbalance was addressed by utilizing cost‑sensitive learning to evaluate the models’ performance using both macro F1 scores and Sharpe ratios adjusted for transaction costs. The results of my analysis show that the best horizon for forecasting is asset specific, where CNNs performed the strongest at horizons ranging from 15 days to 30 days. Additionally, it was found through an ablation study that incorporating sentiment provided a ‘sentiment premium’ for altcoins (+3.28% F1 score for ETH) while Bitcoin remained technically driven. Overall, this risk‑aware model will provide a solid decision‑making platform within the rapidly changing world of digital asset markets.