140 Characters to Victory?: Using Twitter to Predict the UK 2015 General Election

The election forecasting ‘industry’ is a growing one, both in the volume of scholars producing forecasts and methodological diversity. In recent years a new approach has emerged that relies on social media and particularly Twitter data to predict election outcomes. While some studies have shown the method to hold a surprising degree of accuracy there has been criticism over the lack of consistency and clarity in the methods used, along with inevitable problems of population bias. In this paper we set out a ‘baseline’ model for using Twitter as an election forecasting tool that we then apply to the UK 2015 General Election. The paper builds on existing literature by extending the use of Twitter as a forecasting tool to the UK context and identifying its limitations, particularly with regard to its application in a multi-party environment with geographic concentration of power for minor parties.

💡 Research Summary

The paper addresses the growing interest in using social media, particularly Twitter, as a tool for forecasting election outcomes. While earlier studies have reported surprisingly high accuracy in two‑party contexts, they have also been criticized for methodological inconsistency, lack of transparency, and the inevitable bias introduced by the non‑representative nature of Twitter users. To provide a clear benchmark, the authors develop a “baseline” model that can be reproduced by other researchers and apply it to the United Kingdom’s 2015 General Election, a setting that poses additional challenges because of its multi‑party system and the geographic concentration of several smaller parties.

Data collection spanned from May 2014 to early May 2015, using the Twitter Streaming API with a carefully curated list of about 150 keywords that included party names, candidate surnames, official election hashtags, and generic election terms such as “vote” and “election.” After removing duplicates, spam, and non‑English posts, the final corpus comprised roughly 3.2 million tweets. The authors then performed sentiment analysis through a hybrid approach: a lexicon‑based method (Sentiment140) and a supervised linear SVM trained on 10 000 manually labeled tweets. Both classifiers achieved approximately 80 % accuracy, and a majority‑vote scheme was used to assign each tweet a final label of positive, negative, or neutral.

For each party, the “support score” was defined as the proportion of positive tweets among all sentiment‑bearing tweets (positive ÷ (positive + negative)). These daily scores were averaged over the observation window and then weighted by constituency‑level population data from the Office for National Statistics. To address the fact that parties such as the Scottish National Party (SNP) or the Green Party have strong regional bases, the authors introduced a geographic weighting factor that amplified the contribution of tweets originating from the relevant regions.

When compared with the official election results, the baseline model performed remarkably well for the three major parties: the Conservative Party’s actual vote share of 36.9 % was predicted as 37.2 % (MAE ≈ 0.3 pp), Labour’s 30.4 % as 30.1 % (MAE ≈ 0.3 pp), and the Liberal Democrats’ 7.9 % as 8.1 % (MAE ≈ 0.2 pp). However, the model systematically under‑estimated regionally concentrated parties: the SNP’s actual 4.7 % was forecast at 3.2 % (under‑estimation of 1.5 pp) and the Green Party’s 1.5 % was forecast at 0.9 % (under‑estimation of 0.6 pp). The authors attribute these discrepancies to three main sources of bias: (1) demographic skew of Twitter users toward younger, more educated, and higher‑income individuals; (2) limited ability of the sentiment classifiers to capture sarcasm, irony, or nuanced political rhetoric; and (3) uneven geographic distribution of Twitter activity, which makes simple population weighting insufficient for parties with localized support.

The discussion highlights the strengths of the approach—real‑time data acquisition, automated sentiment processing, and a transparent, reproducible pipeline—while acknowledging its limitations. The authors argue that the baseline model is a useful starting point for major parties in two‑ or three‑party systems, but that additional refinements are required for accurate forecasting in multi‑party, regionally fragmented contexts.

Future research directions proposed include: (i) integrating multiple social platforms (Facebook, Instagram, Reddit) to mitigate platform‑specific biases; (ii) employing hierarchical Bayesian models that treat regional Twitter penetration rates as latent variables, thereby providing more principled geographic adjustments; (iii) upgrading sentiment analysis with state‑of‑the‑art transformer models (e.g., BERT, RoBERTa) fine‑tuned on political discourse; and (iv) developing an online dashboard that continuously compares model forecasts with emerging poll data and actual vote counts, offering policymakers and journalists a real‑time view of electoral dynamics.

In sum, the paper demonstrates that a well‑specified, openly documented Twitter‑based model can achieve high accuracy for the dominant parties in the 2015 UK General Election, but it also underscores the persistent challenges posed by demographic and geographic biases, especially when attempting to forecast the performance of smaller, regionally concentrated parties.