Market sentiment and attention driven algorithmic trading strategy

This project investigates whether machine learning models trained on market sentiment and attention-based data can generate profitable directional trading signals for the SPDR S&P 500 ETF (ticker: SPY). Three sources of data are used: Google Search Volume Index data for twenty financial and economic keywords, the CBOE VIX Index, and the CBOE SKEW Index. From these sources, 52 predictive features are engineered and used to train six tree-based classification models (Random Forest, XGBoost, and LightGBM, each in two training window configurations), together with a soft-voting ensemble. All models are evaluated using a walk-forward validation scheme with nested time-series cross-validation and randomized hyperparameter search. Statistical accuracy across models ranges from 50.5% to 52.8%, figures that are statistically close to a naive random walk baseline. Two strategies are backtested — long-flat and long-short — each under zero and realistic transaction costs. Under zero costs, the ensemble long-flat strategy achieves an annualized return of 12.22% and a Sharpe ratio of 0.68, compared to 12.44% and 0.58 for buy-and-hold. Once realistic transaction costs are applied, this risk-adjusted edge disappears and all models fall below the passive benchmark.