Bollen, Mao & Zeng (2011) - Twitter Mood Predicts the Stock Market
Key Insights
- Bollen, Mao and Zeng show that the 'Calm' mood dimension from Twitter feeds can predict DJIA movements with 87.
- 6% accuracy, demonstrating the predictive power of social media sentiment.
Extracted Variables
| Variable | Value | Supporting passage |
|---|---|---|
| Accuracy | 73.3% |
The Findings The model including the Calm mood dimension predicts the direction of DJIA moves with 87.6% accuracy, substantially above the baseline accuracy… |
Edit on GitHub — registry.json
Overview
Bollen, Mao and Zeng (2011) show that public sentiment from Twitter predicts DJIA movements with up to 87.6% accuracy, demonstrating the predictive power of social media sentiment as an alternative data source. The paper became the canonical reference for sentiment-based market prediction.
The Data
The study collects roughly 9.85 million tweets between February and December 2008, scored by two sentiment instruments: OpinionFinder, which gives a binary positive/negative polarity, and GPOMS, a six-dimensional mood analyzer measuring Calm, Alert, Sure, Vital, Kind, and Happy. The DJIA closes over the same period form the prediction target.
The Method
The analysis first measures the mood time series and its lead-lag relationship with the index, then builds a prediction model using a self-organizing fuzzy neural network. The crucial methodological decision is which mood dimensions enter the model: aggregate positive/negative polarity adds little, but the Calm dimension — inversely related to anxiety — carries the predictive signal days ahead of market moves.
The Findings
The model including the Calm mood dimension predicts the direction of DJIA moves with 87.6% accuracy, substantially above the baseline accuracy without mood data (73.3%). The predictive horizon is several days — mood today forecasts the index's direction up to a few days later — which the authors interpret as a sentiment propagation effect rather than same-day noise.
Why It Matters
The paper opened the alternative-data era for sentiment research: if public mood measured from text predicts index moves, then news flow, search volumes, and message boards all become legitimate signals. Its limitations matter equally: a single 2008 sample, one index, and the fragility of social-media measurement mean the 87.6% figure is a proof of concept, not a trading edge to be reproduced blindly.
Key Takeaways
- Polarity is weak; mood dimensions are strong — Calm carried the signal where positive/negative scores added nothing.
- The predictive horizon is multi-day, consistent with gradual sentiment propagation.
- Treat the accuracy as a demonstration, not a deployable edge: sample, target, and measurement all matter.