Improved Stock Market Returns From Systematically Trading Infrequent News

The hypothesis is based on the idea that "hard" news which have a material impact on a company's valuation, such as M&A activity or analyst and credit rating changes, occur less frequently than "soft" news such as technical analysis indications, opinions, and insider transactions.

Further we want to measure the marginal impact of news coverage after an event has been reported. That is, given continued coverage of a company event, how quickly does the impact of similar news articles diminish?

Executive Summary:
In this research, we test the hypothesis that investors in stocks react more strongly to news events that are less frequent. We use RavenPack’s new “event similarity gap measure” introduced in the recently launched RPNA version 4.0 dataset, which measures the number of days since a similar news event was detected. We find the event similarity gap is useful in filtering out recurring, often less impactful news events, specifically:

• Filtering news events with similarity gap greater than 90 days significantly improves long/short strategies for the Russell 2000 and 3000 built on our company sentiment indicator.
• The impact is most profound on the Russell 2000, where we find the Sharpe Ratio improves from 1.03 to 1.43 during the period from Jan 2007 through Aug 2014 (see figure below).
• Similarity gap filtering has a particularly positive contribution to recent performance, particularly 2013/2014, with an improvement of up to 65% in the annual Sharpe Ratio.

Introduction

In this paper we test the hypothesis that investors in companies react more to news events that are infrequently observed. The hypothesis is based on the idea that “hard” news which has a material impact on a company’s valuation, such as M&A activity or analyst and credit rating changes, occurs less frequently than “soft” news such as technical analysis indications, opinions, and insider transactions. Further we want to measure the marginal impact of news coverage after an event has been reported. That is, given continued coverage of a company event, how quickly does the impact of similar news articles diminish?

To test the hypothesis we use a new measure of novelty introduced in RavenPack News Analytics (RPNA) Version 4.0 called the Event Similarity Gap which measures the number of days since a similar event was detected for a company. The concept of similarity gap differs from the traditional novelty measurement in RPNA called Event Novelty Score (ENS) by extending the measurement period from 24 hours to a much longer period of 100 days. The Event Similarity Gap spans more than a typical quarterly earnings cycle addressing an important seasonal component in the news cycle. We believe this measure provides a useful tool to filter out trivial news that adds more noise than information.

Based on RPNA version 4.0, we find evidence that news with a longer similarity gap usually contains more substantial information than news with a shorter similarity gap. For example, novel “analyst-ratings”, “credit-ratings”, and “marketing” news for the same entity tend to be reported very infrequently; while “order-imbalances” and “technical-analysis” news receive much more intensive coverage.

Specifically, filtering news events with a similarity gap greater than 90 days significantly improves our long/short trading strategy built on our company Sentiment Strength Indicator (SSI) for Russell 3000 stocks, and especially for the Russell 2000 stocks. For the trading strategy built using the SSI for on RPNA Full Edition2, the similarity gap filter increases the Sharpe Ratio from 1.03 to 1.43 for Russell 2000 stocks - covering the period from January 2007 through August 2014. More interestingly, we find that the positive contribution to trading performance is mainly observed in recent years including 2013 and 2014.

In the following section, we provide a brief overview of the data used in this paper. In Section 3, we compare the return predictive power of our company sentiment indicators with similarity gap 10, 10 90. Finally, in Section 4 we present our conclusions.