This paper investigates whether Google search query volumes can predict day-ahead stock returns for S&P 100 constituents from 2005 to 2017. Using gradient boosted decision trees, the author models the binary outcome of a stock's return exceeding the index median. Features include cumulative historical returns over 13 horizons, percentage changes in search volume index (SVI) over the past 5 days, and generic information (industry, weekday, trading volume change). Models are trained on rolling 2-year windows and tested on the subsequent year, with 11 non-overlapping test periods. Seven model variants are evaluated, combining different feature sets.

The results show that all feature sets contain predictive information, with average AUC values between 54.2% and 56.7%, significantly above random chance. A simple statistical arbitrage strategy—long the 5 stocks with highest predicted probability of outperformance and short the 5 with lowest—yields annual returns exceeding 57% before transaction costs. The best performing model, using all features, achieves a cumulative return of over 1500% over 11 years, far surpassing the S&P 100 index. The study provides evidence that internet search behavior contains valuable signals for short-term stock return prediction, challenging weak and semi-strong forms of market efficiency. However, transaction costs are not considered, and the practical profitability may be lower.