Stock Price Prediction ML

This research investigates whether incorporating Twitter sentiment analysis improves LSTM-based stock price prediction. The study tested three major stocks — Apple (AAPL), Tesla (TSLA), and Microsoft (MSFT) — comparing baseline technical-indicator models against sentiment-enhanced variants over a one-year period.

Contrary to popular assumptions in financial machine learning literature, the research found that sentiment features consistently degraded prediction accuracy across all three stocks, providing empirical evidence against the naive integration of social media sentiment into price prediction models.

The hypothesis was tested by comparing a one-layer baseline LSTM trained on technical indicators against a three-layer sentiment-augmented LSTM that added daily Twitter sentiment metrics (mean polarity, polarity dispersion, and tweet count). Both models used early stopping and dropout, and were validated through five-fold time series cross-validation preserving chronological ordering. Across all equities, the increase in RMSE was 32.1%, with only Tesla showing statistically significant degradation (t = 6.50, p = 0.003). The sentiment models showed signs of overfitting — smaller training losses but greater validation losses — and permutation importance analysis indicated that sentiment features contributed less than 5% to total predictive importance. These findings suggest that publicly available tweet-level sentiment data may contain insufficient information to improve predictions for highly traded, large-capitalization technology companies, and may instead reduce model performance due to excessive noise.