Financial time series forecasting is traditionally based on historical price patterns, often overlooking exogenous and behavioral variables that can significantly affect market movements. In this study, we enhance the Prophet forecasting framework by integrating sentiment signals derived from financial news headlines and social media platforms such as Twitter. These sentiment indicators, processed using both lexicon-based (TextBlob) and transformer-based (BERT) models, are aligned temporally with stock indices data (notably the Dow Jones Industrial Average) to create a multimodal dataset. The sentiment scores are incorporated into Prophet as external regressors, allowing the model to learn associations between market mood and price trends. This hybrid approach captures both technical and psychological dimensions of financial behavior. Extensive experiments on the DJI index demonstrate that the sentiment-augmented model consistently outperforms the baseline Prophet implementation across all evaluation metrics. For example, the mean absolute percentage error (MAPE) is reduced from 7. 63 to 5. 13% in the test dataset and from 4. 25 to 3. 12% in the validation dataset. These performance gains are particularly evident during periods of high volatility, such as macroeconomic announcements and geopolitical developments, where traditional models tend to reach their limits. By combining quantitative signals with qualitative sentiment, our approach provides a more adaptive and accurate forecasting system. This approach is very promising for traders, portfolio managers, and political analysts seeking to predict market behavior in an increasingly sentiment-driven environment.
Najem et al. (Tue,) studied this question.