Using Sentiment to Create Theme-based Alphas

Executive Summary:

In this paper we address the non-uniformity of stock price reaction to company news sentiment across corporate event types. This inconsistency means that using simple aggregate measures of event-based sentiment in portfolio construction may not be the most effective use of sentiment data. It may be more profitable to treat sentiment thematically. Thus, we create a set of company indicators that look to capture theme-based sentiment from corporate news and evaluate their performance in isolation. We intend to use these indicators in later research to construct profitable stock portfolios.

Highlights of Using Sentiment to Create Theme-based Alphas:

• For most part, market response is consistent with sentiment across a ten day period, particularly for negative news events and small capitalization stocks.
• There is a strong regional overlap for event groups, with the US and Europe having 21 event groups in common for both small and large/mid-cap companies.
• There is also strong overlap amongst important event groups across size and region with 16 event groups ‘surviving’ across both dimensions.
• We were able to identify the ‘strongest’ event groups in terms of market response as seen in the table below.
• There are many idiosyncrasies to market behavior after certain event types, i.e., some event types have strong reversal or momentum effects which may differ by region and/or market capitalization.

Introduction

In this paper we address the fact that reaction of stock prices to company news sentiment is not uniform across corporate event types. This issue means that using simple aggregate measures of sentiment, like moving averages, in portfolio construction may not be the most effective use of sentiment data.

Hence, in this paper we dissect the sentiment signal into event themes, identifying how the sentiment signal decays by event type. With RavenPack’s event taxonomy, we are able to study how the event-based sentiment signals decay in the period following the event. We also break out the results by company size and region. This information could be used when deciding on the directionality of investing and/or holding periods related to the sentiment of particular event types.

This study is similar to the Macquarie Equities Research study “Quantifying Events” (2012) in that we use RavenPack’s event taxonomy to create a set of trading signals. But instead of focusing on granular events, we create a set of daily corporate indicators to measure sentiment at event group level. An event group might be “earnings” or “mergers and acquisitions”, ”analyst-ratings” and so on. This “thematic” approach helps alleviate noise and to better understand the price impact of each theme across various investment horizons; something we use to create a set of “alphas” that could ultimately be combined into a portfolio. By treating each theme as a separate alpha, we hope to identify a set of orthogonal strategies that, in combination, provide a better risk-return trade-off with the additional benefit of lower turnover.

In this paper, however, we only focus on understanding what alphas might add value in our portfolio construction process, while the portfolio construction itself will be addressed in future research.

As we mentioned, the impact of news on stock prices is not always straightforward and we can expect one of three possible scenarios: (1) prices find equilibrium quickly, which results in fast signal decay, (2) prices initially underreact, which leads to prices continuing to move with sentiment, or (3) prices initially overshoot, which leads to a price reversal and hence a move against sentiment.

To evaluate whether prices are typically reversal or momentum driven for given event types, we split each theme-based alpha into a long and short leg. This will help us understand how our indicators should be handled as part of our overall portfolio construction i.e. should we trade sentiment spreads or individual legs, or allow our price prediction to move against sentiment. As a robustness check, we apply our methodology across both company size and region. Specifically, we’ll focus on the Russell 1000 and Russell 2000 for the U.S. as well as their Russell Europe equivalents (i.e. mid/large cap vs. small cap stocks).

In the following section, we provide a brief overview of the data used in this paper. In Section 3, we present our sentiment indicator methodology and in Section 4 we evaluate our indicators as part of an event study - tracking the performance of our positive and negative alphas as well as their spread. Finally, in Section 5 we present our conclusions.