Building a New Predictive Signal from our Quantitative News Feed

Context Analytics is the leader in Unstructured Financial Data and the Quantitative News Feed is one of the fastest-growing products. This feed monitors thousands of securities daily, gathering articles from hundreds of diverse news sources to assess the sentiment surrounding these companies. By pulling in thousands of articles each day, it provides a comprehensive view of how companies are being discussed in the media. The extensive range of sources ensures that we capture a wide array of perspectives, from major news outlets to niche industry publications, giving a well-rounded and accurate measure of sentiment across the market. This breadth of coverage allows us to identify trends, and sentiment shifts that might otherwise go unnoticed, offering deeper insights into how news impacts market movements.

In this blog, we introduce a Weighted Average Sentiment signal, specifically designed to predict short-term returns. The signal is constructed using two key metrics from our data feed: Average Sentiment and Word Count of news articles.

To enhance the accuracy and relevance of our signal, we further refine the dataset by incorporating two additional fields: Relevancy and Reposted. The Relevancy field, ranging from 0 to 100, indicates how pertinent an article is to a specific security, while the Reposted flag (0 or 1) identifies whether the content of the article has already been published by other outlets. For this analysis, we focused exclusively on articles with a Relevancy score of 100 and a Reposted flag of 0, ensuring that the news was distinct and highly relevant to each security.

Moreover, we narrowed our dataset to include only those securities with a 20-day z-score of volume greater than 0—indicating that these securities were receiving more media attention than usual. This filtering process allowed us to focus on securities that were prominently featured in the news.

Constructing the Weighted Average Sentiment Metric

The Average Sentiment field represents the tone of an article—whether it’s positive, negative, or neutral. The Word Count measures the length of the article, capturing the amount of detail given to a particular topic.

To calculate our metric for each ticker on a given day, we sum the Average Sentiment of all articles related to that ticker, each weighted by its Word Count. We then normalize this sum by dividing it by the total Word Count for all articles about that ticker on that day. Here’s how it works:

1. Weighted Sum of Sentiment: For each article mentioning a ticker, multiply the Average Sentiment by its Word Count. This step ensures that longer articles have a greater influence on the overall sentiment score.
2. Normalization: Sum the products from the previous step across all articles for each company, and then divide by the total Word Count across all articles for each company. This normalization balances the sentiment across all articles, providing a holistic view of the sentiment landscape.

Mathematically, this can be written as the following:

The rationale behind this approach: longer articles—those with more words—tend to delve deeper into a topic, offering more nuanced perspectives. As such, they should naturally have a greater impact on the overall sentiment score. For instance, if a long article has a strong positive average sentiment, it will pull the weighted average sentiment upwards more significantly than a shorter article with the same average sentiment.

Practical Impact on Market Reactions

The Weighted Average Sentiment metric provides clear insights:

- Positive Reaction: Higher scores indicate more positive news coverage, as the sentiment reflects favorable analysis across multiple sources. This can potentially lead to stronger positive market reactions.

- Negative Reaction: Lower scores signify more negative news coverage, as the sentiment captures a negative tone in the articles. This may result in more pronounced negative market reactions.

Testing the Predictive Power

To test the predictive nature of this metric, we analyzed a universe of securities with a price greater than $5. Each day, before market close, we calculated the Weighted Average Sentiment for the previous 24 hours and grouped the securities into quintiles. Positions were held for the subsequent close-to-close period, and quintiles are recalibrated daily. Securities within each quintile are equally weighted.

Since August 2021, we have observed a strong correlation between the Weighted Average Sentiment and subsequent close-to-close returns. Securities in the top quintile—those with the most positive news—outperformed the others, while those in the bottom quintile underperformed. The quintiles fell into a monotonic spread, indicating that this could be a predictive metric.

Our Quintile 5 - Quintile 1 long-short strategy significantly outperforms, delivering superior Sharpe and Sortino ratios. This suggests that the metric effectively captures both the positive and negative impacts of sentiment on companies.

Unlocking Predictive Metrics with Our News Feed

The Weighted Average Sentiment signal is just one example of how our Quantitative News Feed can be harnessed to construct predictive metrics. With countless possibilities for customization, you can tailor these insights to fit your specific investment strategies.

For more information or to start a trial with this dataset, visit www.contextanalytics-ai.com.