Unstructured text in newspaper articles can potentially capture investor sentiment or biases and is valuable in predicting future stock market returns. By quantifying these text data and combining them with familiar regression factors, the authors find that future market returns can be better predicted. Predictive power is further improved with cluster analysis.
What’s Inside?
The authors extract information from newspaper articles and aim to discern whether quantitative measures of these data are useful in predicting future stock market returns. They examine the content of news, quantify the data using word-count indices, and investigate whether the indices are useful in predicting future returns of the German DAX index. They also examine the effect of clustering on predictive power and explore whether the predictive power of news articles has increased over time.
How Is This Research Useful to Practitioners?
Studying the German stock market, the authors find that financial newspaper articles contain information that is valuable in predicting future stock returns. News stories have statistically significant predictive potential in and out of sample, even after controlling for such commonly known factors in regression models as size, value, volatility, term spread, and so forth. The predictive potential of newspaper content has increased over time, according to the authors’ out-of-sample analysis.
They find that word-count indices are significantly correlated with future stock returns. Several words are Granger causative of future DAX returns. The use of statistical word clusters increases predictive power. Seven or more clusters are essential for optimally predicting returns; fewer clusters (e.g., two clusters with “positive” and “negative” sentiment) reduce explanatory power.
When the authors extend their study to the real economy, they find that predictive ability goes down—suggestive of, perhaps, a behavioral explanation. As an application, the authors suggest that news data can help predict future stock returns in Germany. Asset managers with exposure to the German stock markets may find the study beneficial. Quantitative managers and algorithmic traders may also find the study useful in complementing existing models.
Some may consider that the study is an example of data mining—finding clusters of words that might, ex ante, explain market movements. Further validation via application of the authors’ techniques to other news sources and other markets could help reinforce their contentions. If the explanation for the authors’ results is behavioral, then further clarification as to why this might be would be helpful.
How Did the Authors Conduct This Research?
Unstructured newspaper text forms the authors’ raw research material. From newspaper articles published in 1989–2011 in the Handelsblatt, a German financial newspaper, the authors assemble a 236-word list containing nouns, adjectives, verbs, and words with positive and negative sentiment.
To quantify the text, the authors construct word-count indices by counting monthly word occurrence in articles and standardizing with Z-scores. With these foundational word-count indices, the authors conduct univariate, multivariate, and cluster analyses.
For their univariate analysis, the authors calculate correlations between word-count indices and future-period DAX excess returns and run Granger-causality tests. Granger causality is based on cause preceding effect and the cause evidencing distinct predictive power on the effect’s future value.
The authors conduct a multivariate stepwise regression to assess the predictive ability of content using word-count indices as the dependent variable. They control for such predictors as Fama–French size and value factors, macroeconomic factors, foreign stock market effects, the consumer price index, and volatility; they also control for multicollinearity. Goodness of fit is tested using in-sample analyses, whereas predictive ability is tested using out-of-sample tests with rolling time windows. The regression slope of realized and forecasted returns and root mean square error help evaluate predictive ability.
Cluster analysis is used to reduce high data dimensionality. Cluster analysis is unsupervised learning, where data are grouped based on similarities. This process is not the same as categorizing data into such predefined categories as negative or positive. A hierarchical clustering algorithm determines cluster proximity using complete linkage, where the maximum Euclidean distance between two points in different clusters determines proximity.
Abstractor’s Viewpoint
Advances in knowledge, increased computing power, and large amounts of data may result in more interest in quantifying information from text. This study is based on the German stock market. Future paths to explore include other markets and techniques, as well as potential limitations.