Sentiment-enhanced rice price forecasting under sparse social-media coverage: Evidence from Saudi rice imports

This study examines whether carefully lagged social-media sentiment adds incremental forecast information to Saudi rice import prices once temperature-based climate variables and historical price dynamics are already included. We construct a monthly series for premium Basmati and standard Maza rice using customs-cleared price data, station-based temperature measures, and Arabic Twitter content labeled for sentiment with GPT-4. Because the retained tweet signal is sparse after relevance and engagement filtering, forecasts are evaluated in a constrained monthly setting using seasonal autoregressive integrated moving average with exogenous regressors (SARIMAX), linear regression (ordinary least squares), ridge regression, gradient boosting, and random forest under a time-ordered validation design with a held-out test period. Sentiment features improve forecast accuracy most consistently for Basmati, with smaller gains for Maza; the strongest overall results are achieved by transparent linear models and sentiment-enhanced SARIMAX rather than by the more complex tree ensembles. The findings indicate that large language model-assisted sentiment labeling can provide useful demand-side information when paired with disciplined feature engineering, while the magnitude of the benefit depends on commodity type, data coverage, and market structure.