This study thoroughly integrated news information into the daily stock price up/down prediction. The predictors included the carefully selected technical features of stock prices, the word level and sentence level sentiment features, and the automatically defined polarity scores of news articles. The main goals of this study are to establish the functional polarity scores of news articles, build models with high prediction accuracy, and identify the important news feature effect for the price up/down prediction. Compared with the Long Short-Term Memory (LSTM) model of Recurrent Neural Networks, we take advantage of Multivariate Adaptive Regression Splines (MARS) as our primary prediction model for its capability in local feature building and its interpretability of selected features. An empirical analysis was carried out on the selected 100 individual stocks in Standard & Poor’s 500 from 2019 to 2021. The overall average prediction accuracy for 100 stocks by the proposed MARS model was 0.653 which was significantly higher than 0.505 by the LSTM model.