Abstract

This study shows the current potential of bilingual sentiment analysis as a predictive tool for forecasting Saudi equity returns over a one-week horizon. Dataset is a combination of daily observations for 279 publicly listed companies and sentiment indicators based on nearly 19,300 financial news articles. The sentiment indicators were assessed using advanced NLP models, namely FinBERT for English and AraBERT for Arabic, and subsequently aggregated daily per firm. To forecast the five-day relative returns, three of the most sophisticated learning models, i.e., LSTM, GRU, and 1D-CNN, were trained and evaluated in a walk-forward validation framework. The enhanced ensemble model reduced the RMSE to 0.0328 and the MAE to 0.0224, compared with the baseline model’s RMSE of 0.0342 and MAE of 0.0238. This represents a 25% to 30% reduction in predictive error, in addition to an improvement in directional predictive accuracy from 0.55 to 0.78. 

Statistics