Abstract
The fast growth of mobile technologies has transformed social media, particularly Twitter, into a vital platform for individuals to share their opinions, emotions, and sentiments in real time. Understanding these public sentiments offers valuable insights for financial market forecasting. This study investigates the predictability of asset return dynamics using text data extracted from Twitter from April 3, 2023, to July 28, 2023. We analyse posts related to conventional assets (S&P 500, NASDAQ, and Gold) and digital assets (Bitcoin and Ethereum) to determine the sentiment embedded in the text and use it as an explanatory variable to forecast asset return movements. The study contributes to the literature by integrating sentiment analysis into return forecasting models and by comparing predictive performance across traditional and cryptocurrency markets. To this end, we first computed sentiment scores using the Valence Aware Dictionary and Sentiment Reasoner (VADER). Second, we developed an ensemble machine learning model to capture and analyse sentiment using a hard voting mechanism to integrate Logistic Regression, Naïve Bayes, and Support Vector Machine classifiers. Third, the sentiment scores were incorporated as exogenous variables into the Autoregressive Integrated Moving Average with Exogenous Variables (ARIMAX) model to predict asset returns. To evaluate the effectiveness of the ARIMAX models, their prediction performance was benchmarked against two deep learning models: Long Short-Term Memory (LSTM) and Bidirectional Long Short-Term Memory (BiLSTM). Furthermore, we evaluate return prediction accuracy using several metrics, including Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE). Finally, the Diebold-Mariano (DM) test was used to statistically compare the predictive accuracy of the forecasting models. The empirical results show that investor sentiment derived from X posts carries significant predictive power for asset returns. The deep learning models slightly outperformed the ARIMAX model; however, the differences were exceedingly small. Furthermore, the Diebold-Mariano test revealed no statistically significant difference between ARIMAX and the deep learning models for most assets, except Bitcoin, where BiLSTM significantly outperformed ARIMAX. This outcome suggests that ARIMAX remains competitive with deep learning models despite their slightly higher predictive accuracy.