Key points are not available for this paper at this time.
Abstract Macroeconomic forecasting is critical for economic stability and informed policy decisions, yet conventional approaches often rely on delayed structured data and ignore real-time public sentiment. This study proposes a deep learning (DL) framework integrating social media sentiment with traditional macroeconomic indicators to evaluate whether sentiment signals can enhance prediction in a data-constrained setting. Using 4000 time-stamped public posts containing economy-related keywords, text preprocessing includes duplicate filtering, normalization, and lemmatization to reduce noise and linguistic variability. Word2Vec embeddings capture semantic context, while transformer-based models compute sentiment polarity and subjectivity scores. These signals are temporally aggregated and aligned with macroeconomic indicators, assuming that aggregated sentiment may act as a leading informational signal. A Seven-Spot Ladybird-optimized Stacked Long Short-Term Memory (SSL-Stacked LSTM) model is applied to capture multi-level temporal dependencies between sentiment and macroeconomic variables. Experiments conducted in Python achieve an R2 of 0.9428 and an MAE of 0.0246, indicating promising predictive performance. Overall, findings suggest that social media sentiment can complement traditional indicators in macroeconomic forecasting, while broader validation across diverse datasets and temporal conditions is required to establish reliability and applicability.
Geng et al. (Sun,) studied this question.