https://link.springer.com/article/10.1007/s40745-025-00666-0

Trading decision-making is significantly influenced by psychological resistance that emerges under dynamic market conditions. Fear and greed states provide a quantifiable representation of these behavioral dynamics, serving as the basis for sentiment modeling. This study proposes a novel set of buy-and-sell pressure induced lagged features and integrates them with machine learning to predict multiclass fear–greed states in the Bitcoin market. To address severe class imbalance across five sentiment categories, we employ a one-step-ahead rolling window backtesting procedure. The predictive performance of extreme gradient boosting (XGBoost), support vector regression (SVR), and long short-term memory (LSTM) networks is systematically evaluated. Results demonstrate that SVR combined with the proposed lagged features achieves the highest performance, yielding an area under the curve (AUC) of 0.93 and outperforming both XGBoost and LSTM. These findings underscore the effectiveness of feature engineering based on buy-and-sell pressure in enhancing sentiment forecasting for volatile cryptocurrency markets. Beyond predictive accuracy, the framework offers practical applicability, enabling data-driven trading strategies and integration into automated trading systems for continuous market monitoring and decision execution.