Source-linked AI summary
The Effects of Twitter Sentiment on Stock Price Returns
Gabriele Ranco, Darko Aleksovski, Guido Caldarelli, Miha Grčar, Igor Mozetič
TL;DR
The paper asks whether Twitter sentiment about DJIA companies is related to stock returns despite limited full-period dependence. It adapts event-study analysis to Twitter volume peaks and finds significant sentiment–return dependence, including around less obvious events. The study also identifies scope boundaries requiring future tests of forecasting power, finer time resolution, broader company coverage, and longer periods.
Problem
Limited empirical dependence between financial time series and web-derived series motivates testing whether Twitter sentiment relates to stock returns around selected activity peaks.
Method
The paper applies an event-study procedure to Twitter volume peaks, using sentiment polarity in tweets about 30 DJIA companies and relating it to subsequent stock returns.
Results
Statistically significant dependence between Twitter sentiment and stock returns appears around Twitter volume peaks, including peaks not tied to expected news.
Takeaways & Limitations
Twitter sentiment around activity peaks provides evidence relevant to understanding or now-casting market behavior, provided data sources and events are properly selected.
Takeaways & Limitations
The study does not yet test forecasting power and is limited to 30 DJIA stocks over a 15-month period, with finer time scales and broader samples left for future work.
Abstract
from arXiv · showhide
Social media are increasingly reflecting and influencing behavior of other complex systems. In this paper we investigate the relations between a well-know micro-blogging platform Twitter and financial markets. In particular, we consider, in a period of 15 months, the Twitter volume and sentiment about the 30 stock companies that form the Dow Jones Industrial Average (DJIA) index. We find a relatively low Pearson correlation and Granger causality between the corresponding time series over the entire time period. However, we find a significant dependence between the Twitter sentiment and abnormal returns during the peaks of Twitter volume. This is valid not only for the expected Twitter volume peaks (e.g., quarterly announcements), but also for peaks corresponding to less obvious events. We formalize the procedure by adapting the well-known "event study" from economics and finance to the analysis of Twitter data. The procedure allows to automatically identify events as Twitter volume peaks, to compute the prevailing sentiment (positive or negative) expressed in tweets at these peaks, and finally to apply the "event study" methodology to relate them to stock returns. We show that sentiment polarity of Twitter peaks implies the direction of cumulative abnormal returns. The amount of cumulative abnormal returns is relatively low (about 1-2%), but the dependence is statistically significant for several days after the events.
Introduction
The paper examines whether Twitter activity and sentiment about DJIA companies relate to stock-market behavior, focusing on short periods around activity peaks rather than only full-period correlations. It adapts event-study methods to show statistically significant relations between Twitter sentiment and stock returns around such events.
- Motivation: Social-network data may help study collective investor behavior because herding and panic can contribute to financial contagion and crises.The authors connect monitoring anomalous collective behavior with possible policy-maker interventions.
- State-of-the-art: Existing work finds relations between Twitter activity or mood and financial markets, but empirical correlations between web-derived and financial time series remain limited.Prior studies examine tweet volume, sentiment, stock prices, trading volume, and DJIA mood indicators.
- Approach: The study analyzes sentiment in tweets about 30 DJIA companies using daily sentiment time series, Pearson correlation, Granger causality, and event-study analysis.The event-study approach focuses on shorter periods around Twitter activity events and uses aggregate tweet sentiment rather than earnings-announcement sentiment.
- Findings: Restricting analysis to shorter periods around Twitter events reveals a statistically significant relation between Twitter sentiment and stock returns.The result is consistent with existing literature on the information content of earnings.
- Findings: Two independent studies using event-study adaptations for Twitter data corroborate statistically significant links between financial Twitter data and stock returns.The studies use different stocks, periods, sentiment classifiers, event-detection algorithms, and significance tests, while reaching similar conclusions.
- Contributions: The released Twitter sentiment time series support studies of alternative sentiment aggregations, events, and post-event effects such as drifts, reversals, and volatility changes.The paper also contributes a high-quality sentiment classifier trained on over 100,000 annotated tweets and evaluated against financial-expert agreement.
Data
The dataset combines daily stock returns for 30 DJIA companies over 15 months with Twitter data collected through cash-tag searches and sentiment labels generated by an SVM classifier. It provides tweet-volume and sentiment measures for subsequent market analysis.
- Market data: The market sample contains 30 DJIA stocks observed over 15 months between 2013 and 2014.The investigated ticker list is provided in Table 1.
- Market data: Daily stock returns are computed from closing prices using raw returns rather than log returns.The choice follows the original event-study methodology.
- Twitter data: Twitter data consist of relevant cash-tagged tweets collected through the Twitter Search API, with sentiment assigned to the complete tweet set.The paper states that available tweets matching the cash-tag queries were acquired.
- Sentiment data: Over 100,000 tweets were labeled by 10 financial experts with negative, neutral, or positive sentiment labels, and an SVM model classified over 1.5 million tweets.The resulting data are organized as sentiment time series.
- Twitter measures: The Twitter measures include daily total, negative, neutral, and positive tweet counts, plus sentiment polarity based on positive and negative tweets among non-neutral tweets.The sentiment time series are publicly available for the DJIA 30 stocks.
- Twitter data: Table 1 records the collected Twitter data for the 15-month period by company name and tweet count.The table is associated with the collected company-level Twitter sample.
Methods
The paper combines sentiment classification, time-series tests, and an event-study framework to examine whether Twitter activity and sentiment relate to stock returns. It detects Twitter-volume peaks, classifies their polarity, and evaluates abnormal returns around earnings-announcement and other events.
- Sentiment classification: The study classifies tweets into negative, neutral, and positive sentiment using supervised machine learning with manually annotated data, SVM classifiers, cross-validation, and application to all collected tweets.The two-classifier ordinal SVM approach partitions tweets into three sentiment areas; disagreements are labeled neutral.
- Sentiment classification: Twitter preprocessing removes URLs, cash-tags, and user mentions, collapses repeated letters, and applies tokenization, lemmatization, n-grams, and TF-IDF weighting.Stop words are retained because removing terms such as “not” can alter sentiment polarity.
- Time-series analysis: Pearson correlation measures contemporaneous linear dependence, while Granger causality tests whether Twitter variables improve prediction of stock returns through stationarity checks, VAR modeling, residual tests, and an F-test.The Granger procedure uses the ADF test, selects VAR order with AIC, BIC, FPE, and HQIC, and compares baseline and extended models.
- Event study: The event-study framework identifies abnormal events, groups them by polarity, and compares actual returns with expected normal returns over a defined event window.The paper uses a market model in which expected stock returns depend linearly on DJIA returns, with abnormal returns defined as the residual difference.
- Event detection: The procedure detects 260 peaks, including 118 of 151 earnings-announcement events, corresponding to 78% recall, while also identifying non-earnings-announcement events.The analysis examines both known earnings-announcement effects and less obvious events whose price impact is not established in the cited passage.
Results
Across the full period, correlation and Granger-causality evidence is limited, but event-study analysis finds statistically significant links between Twitter sentiment peaks and stock returns. The effect is directionally consistent and remains after excluding earnings announcements.
- Twitter sentiment classification: The sentiment classifier reaches annotator agreement across all three evaluation measures using 10-fold cross-validation.The classifier was evaluated on 103,262 annotated tweets, while the related comparison used 2,500 examples and achieved 64.2% Accuracy.
- Correlation and Granger causality: Only three companies pass the Granger test for sentiment polarity predicting price return.This indicates that polarity is generally not useful for predicting price return over the full period.
- Correlation and Granger causality: For one third of companies, tweet volume Granger-causes absolute price return, indicating predictive information about price volatility.This extends earlier aggregate-index findings to individual stocks.
- Event study: The event study finds statistically significant relations between Twitter sentiment and stock returns around shorter event windows.The analysis applies event-study methodology after examining full-period correlation and Granger causality.
- Cumulative abnormal returns: After positive Twitter peaks, average CAR increases; after negative peaks, it decreases, with positive-event CAR significant at the 1% level for ten days.Negative-event losses are twice as large in absolute terms, while neutral-event CAR is very low and mostly insignificant.
- Cumulative abnormal returns: After removing earnings announcements, CAR remains significant at the 1% level for four days after positive events and eight days after negative events.The non-EA effects are smaller and shorter-lived, but remain statistically significant.
Discussion
The paper argues that Twitter sentiment is informative mainly around activity peaks, including unexpected events, rather than uniformly across the full period. It also identifies open directions involving event classification, forecasting, time resolution, and broader coverage.
- Discussion: Twitter activity peaks identify moments when sentiment and market evolution are strongly connected.The authors frame event detection as a way to capture interactions that may be missed over the entire time period.
- Discussion: Similar dependence holds for peaks without expected news about the traded stock.This extends the result beyond anticipated events such as earnings announcements.
- Discussion: The event statistic uses significance thresholds of |θ̂| > 2.58 at 1% and |θ̂| > 1.96 at 5%.Table 4 marks results meeting these thresholds with double and single asterisks, respectively.
- Discussion: Independent work corroborated the findings and classified non-EA events into 16 company-specific categories.That study trained a Naive Bayesian classifier on 2,500 manually classified tweets.
- Discussion: Forecasting power remains an open question, although the proposed application would trade on tweet-volume peaks and matching polarity.The paper also identifies panic-diffusion detection and mitigation as a possible application.
- Discussion: Event detection should use only current and past Twitter volume, while future work may add hourly resolution, more companies, and longer periods.These extensions are presented as planned directions rather than established results.
Supporting Information
The supporting information documents detected events and their polarity, including both earnings-announcement and non-earnings-announcement events, with an associated figure supplied in PNG format.
- Supporting Information: The appendix provides detailed information about detected event dates and their Twitter sentiment polarity.It is labeled “Event dates and polarity.”
- Supporting Information: The appendix reports 118 detected EA events and 182 detected non-EA events.
- Supporting Information: The EAtwitterpeaks figure is available in PNG format.