Can social-media activity add useful information to traditional polling? This paper explores that question by combining opinion polls with Twitter hashtag activity and network structure.
Designed to measure voter sentiment more systematically, but they arrive as periodic snapshots rather than a continuous stream.
Available in real time and rich in campaign dynamics, but noisy, selective, and not representative of the full electorate.
Twitter Hashtag-based Opinion Survey. THOS uses the frequency of campaign-related hashtags over time together with polling information.
In the paper, THOS is used as the simpler model and was most effective in examples with a clearer separation between the two sides.
Twitter Hashtag and Network-based Opinion Survey. THANOS adds network structure to the hashtag-and-poll framework.
The network-enhanced model was developed for closer races, where hashtag counts alone may not capture enough of the campaign structure.
A simple tweet count does not tell the whole story. Two campaigns can generate similar volumes of tweets while having very different patterns of influence.
In the paper, the network is built from retweets and mentions. Four leading influencers anchor the illustrated network—two from each campaign—and the surrounding clusters expand from Day 1 to Day 45 as engagement increases.
High-impact accounts sit near the centers of the major retweet clusters.
The clusters become denser and begin to bridge as information spreads across the campaign network.
Harmonic centrality and retweet proportions summarize network influence that hashtag volume alone cannot capture.
Social-media datasets are incomplete. Tweets can be deleted, accounts can disappear, and historical samples may omit users.
The paper randomly removed up to 30% of tweets and reported that THANOS forecasts changed little. The authors attribute this stability to the emphasis on common hashtags, retweet structure, and highly influential nodes.
The paper’s conclusion is not that Twitter should replace polling. Rather, it presents a hybrid forecasting framework in which traditional polls provide a structured measure of public opinion while social-media activity contributes fast-moving information about campaign attention and network dynamics.
That framing is important because Twitter users are not a representative sample of voters. The contribution is the attempt to use the platform’s dynamics as an additional predictive signal rather than treating raw tweet volume as a direct proxy for the electorate.
Dhrubajyoti Ghosh, William A. Boettcher, Rob Johnston, and Soumendra Lahiri · Data Science in Science, 2025.
This page is an accessible discussion of the published study. Numerical results and model descriptions above are reported from the paper; they are not forecasts of any current election.