Back to Community
Event Study - Trading on Negative News & Blog Article Sentiment

10/22 Update: Join us and Kumesh from Accern as we walk through this research & discuss how to use negative news sentiment in trading strategies in this recorded webinar - http://bit.ly/1OL3iT5

10/16 Update: The first, original version of this research had a few errors. This is an updated version with the correct findings.

A few years back, Paul Tetlock released his influential paper More Than Words: Quantifying Language to Measure Firms' Fundamentals. One of the most startling finds from his research was that negative words in articles captured "otherwise hard-to-quantify aspects of firms' fundamentals." Even more surprising was that "that stock market prices incorporate the information embedded in negative words with a slight delay." In other words, by capitalizing on this delay, there is potential for profit in trading strategies. That being said, the markets are efficient, and much of the information found in news articles are priced in previous to the release of any news piece.

Event Study

In the graph above, you can see that much of the alpha has been squeezed out previous to the release of a WSJ article with a succeeding negative drift in price. However, what if you looked at non-mainstream news sources?

There are thousands and thousands of news sites & blogs that aren't followed by the general public so what if we followed the return streams following these negative news sentiments?

Accern gathers and analyzes almost every blog and news source available on the web. They provide both

  • An article sentiment score (a pure estimate of how positive or negative the article) using NLP and ML
  • An impact score (the probability that a stock's price will change by more than 1% on the next trading day).

They also created multiple trading strategies using both positive and negative sentiment as trading signals. Interestingly, it looks like negative news sentiment can signal a reversal in a stock's price. Take a look at my notebook for more information including a quick summary for anyone that wants a quick skim.

I encourage you to take this research and expand upon it by filtering down securities by sector exposure or using different sentiment score limits (only negative ones are used in this study - explained in the notebook).

As always, we're looking for feedback on how we can make research like this better for you guys. Post below what you'd like to see in further notebooks and backtests.

You can get access to the data now by either cloning the notebook or going here: https://www.quantopian.com/store/accern/alphaone

Loading notebook preview...
Notebook previews are currently unavailable.
Disclaimer

The material on this website is provided for informational purposes only and does not constitute an offer to sell, a solicitation to buy, or a recommendation or endorsement for any security or strategy, nor does it constitute an offer to provide investment advisory services by Quantopian. In addition, the material offers no opinion with respect to the suitability of any security or specific investment. No information contained herein should be regarded as a suggestion to engage in or refrain from any investment-related course of action as none of Quantopian nor any of its affiliates is undertaking to provide investment advice, act as an adviser to any plan or entity subject to the Employee Retirement Income Security Act of 1974, as amended, individual retirement account or individual retirement annuity, or give advice in a fiduciary capacity with respect to the materials presented herein. If you are an individual retirement or other investor, contact your financial advisor or other fiduciary unrelated to Quantopian about whether any given investment idea, strategy, product or service described herein may be appropriate for your circumstances. All investments involve risk, including loss of principal. Quantopian makes no guarantees as to the accuracy or completeness of the views expressed in the website. The views are subject to change, and may have become unreliable for various reasons, including changes in market conditions or economic circumstances.

22 responses

Am I reading it correctly, that the mean drift is -0.01 with a standard deviation of +/- 0.1?

Simon,

Yeah, since the sample size is so large it gets hard to conclude a single, concrete result. It's common in doing event studies - there are differences in volatility even year-over-year (e.g. 2012 versus 2014).

My guess is that the standard deviation becomes much more contained as you, say, limit the study to stocks within a specific sector and so forth.

What are your thoughts on that?

Seong

I'll believe it when I see it :)

Good point.

I definitely encourage people to expand, disprove, dissect the research. I plan on tackling Simon's questions soon as well.

The alphaone_abnormal_returns DataFrame has a colum for each day in the time-frame. Does that mean there was one (and only one) data point each day?

Disclaimer

The material on this website is provided for informational purposes only and does not constitute an offer to sell, a solicitation to buy, or a recommendation or endorsement for any security or strategy, nor does it constitute an offer to provide investment advisory services by Quantopian. In addition, the material offers no opinion with respect to the suitability of any security or specific investment. No information contained herein should be regarded as a suggestion to engage in or refrain from any investment-related course of action as none of Quantopian nor any of its affiliates is undertaking to provide investment advice, act as an adviser to any plan or entity subject to the Employee Retirement Income Security Act of 1974, as amended, individual retirement account or individual retirement annuity, or give advice in a fiduciary capacity with respect to the materials presented herein. If you are an individual retirement or other investor, contact your financial advisor or other fiduciary unrelated to Quantopian about whether any given investment idea, strategy, product or service described herein may be appropriate for your circumstances. All investments involve risk, including loss of principal. Quantopian makes no guarantees as to the accuracy or completeness of the views expressed in the website. The views are subject to change, and may have become unreliable for various reasons, including changes in market conditions or economic circumstances.

I added fundamentals filtering but getting a starange error, any idea ? :

ValueError: Bad response: 500 Internal Server Error
Internal Server Error
The server encountered an internal error and was unable to complete your request. Either the server is overloaded or there is an error in the application.

Loading notebook preview...
Notebook previews are currently unavailable.

joe, I'm unable to duplicate the behavior right now. I cloned the notebook and ran it successfully. I suspect you just got unlucky with a bad connection or some temporary blip. We didn't see anything in our logs or utilization patterns that showed something systematically wrong (as would be suggested by that error message).

But if it continues to happen, let us know or submit a ticket to Help.

Thanks,
Josh

Disclaimer

The material on this website is provided for informational purposes only and does not constitute an offer to sell, a solicitation to buy, or a recommendation or endorsement for any security or strategy, nor does it constitute an offer to provide investment advisory services by Quantopian. In addition, the material offers no opinion with respect to the suitability of any security or specific investment. No information contained herein should be regarded as a suggestion to engage in or refrain from any investment-related course of action as none of Quantopian nor any of its affiliates is undertaking to provide investment advice, act as an adviser to any plan or entity subject to the Employee Retirement Income Security Act of 1974, as amended, individual retirement account or individual retirement annuity, or give advice in a fiduciary capacity with respect to the materials presented herein. If you are an individual retirement or other investor, contact your financial advisor or other fiduciary unrelated to Quantopian about whether any given investment idea, strategy, product or service described herein may be appropriate for your circumstances. All investments involve risk, including loss of principal. Quantopian makes no guarantees as to the accuracy or completeness of the views expressed in the website. The views are subject to change, and may have become unreliable for various reasons, including changes in market conditions or economic circumstances.

Thanks for the questions guys.

Thomas - looks like even when there multiple events per day, we only ended up looking at one event per day. That was an error in the code. We're fixing it now and it looks like the results have changed with the fix. I'll update with new information soon. Thanks for waiting!

For those interested in learning more about news sentiment - how it's built, ways to use it in trading strategies, and more - join our webinar next Thurs. Oct 22 at 1PM EST. We'll be joined by Kumesh from Accern and Josh Payne from Quantopian.

You can register for the webinar here: https://attendee.gotowebinar.com/register/5430134256751399938

Seong

Hey all,

So Thomas pointed out a pretty glaring flaw in the original event study. It was only looking at the last article sentiment for a given date versus the average of all sentiments for a given date. I've fixed that and updated the notebook accordingly.

Long-story short, it looks like negative sentiment articles could signal a stock's price reversal. Please take a look at the attached notebook. Looking forward to your feedback.

Loading notebook preview...
Notebook previews are currently unavailable.

Now that looks much more interesting! Did the standard deviation decrease?

Seong,

Is the asof_date the date of the negative news release? If so, how do you explain the decline in returns starting 5 days prior? Also, are your days calendar days, or trading days?

Grant

How would you explain the fact that when running the notebook with the Event free data , then the results are totally different (e.g. returns go up after the events).
In addition I noticed that the notebook brings positive sentiments although (e.g. for INTEL as example) although the filter is for negative only ?

sorry if I am missing something here.

Loading notebook preview...
Notebook previews are currently unavailable.

Thanks for the questions. We're hosting a webinar this Thursday, October 22 at 1PM EST and if you'd like your questions answered in person, please register here: https://attendee.gotowebinar.com/register/5430134256751399938

Joe,

In the original notebook I also saw that returns went up after the event - which matches what you found in your research. In this case though, we're looking at the ENTIRE dataset versus the limited set that comes with the unpaid. I also chose only negative sentiment news articles because Paul Tetlock's paper stated "that stock market prices incorporate the information embedded in negative words with a slight delay." I sought to see if there was a similar behavior with Accern's sentiment data. It doesn't look like there was - rather, there it was more than people generally overreacted to negative news which causes the reversal in price/increase in return stream after the event.

Grant,

I'm looking at trading days, not calendar days - and the decrease in price you see can be explained by a few factors. (1) The asof_date is the release of the sentiment score versus the article. However, the article release time and sentiment release time could happen on the same day. (2) The decrease in price could be related to the fact that information is readily priced in prior to the release of a news article (i.e. asymmetric access to information). I'm still learning about this and would love your thoughts on this.

Thomas,

Here's a version of the notebook that centers all returns around t=0. This way I got the standard deviation of post-event returns starting from after the event date. It looks like the standard deviation did decrease from the previous version of the notebook - hovering around +/- 7.0% (still quite high) on an approximate average drift of ~ + 1.0%. I think it's worth noting that the previous notebook was looking at cumulative standard deviation from t-20, versus t=0 here.

You can see the std deviation attached here.

Loading notebook preview...
Notebook previews are currently unavailable.

Advice:
Change...

  # To try out the free sample  
from quantopian.interactive.data.accern import alphaone_free



# To try out the free sample  
from quantopian.interactive.data.accern import alphaone_free as alphaone

Keeps the rest of code intact

We just finished up the webinar on this research. For those who want to see us and Kumesh from Accern walk through this research you can view the recorded session here: https://www.youtube.com/watch?v=GlEV3WtquK4&feature=youtu.be

Hi Seong,

The asof_date is the release of the sentiment score versus the article.

Is the asof_date the new release date or the date when Accern published the sentiment assessment? Is there any reason to think that the news release and sentiment dates would be different?

From your plot above, it appears that non-public negative news takes 5 days to be absorbed by the market, at which time (t = 0), the negative news is reported, and then the returns recover? This still doesn't make sense.

Or are you saying that the public announcement, along with the sentiment publication are coincident with the onset of the returns decline (t = -5), and then by t = 0, the returns recover and start going up again (basically, an over-reaction to bad news that has a 5-day lifetime)?

In other words, how should I interpret t = 0? Naturally, one would think it is the time at which negative news & sentiment were published, but is it?

Hi Grant,

Thanks for asking. For this alphaone dataset, the asof_date is when Accern published the rolled-up sentiment assessment. So for example, let's say two articles about AAPL came out on October 1st, 2014. The first at 9:45AM and the second at 10:45AM. A single datapoint would exist for AAPL on 10/1/2014 which would be a rolled-up analysis of both the articles.

Or are you saying that the public announcement, along with the sentiment publication are coincident with the onset of the returns decline (t = -5), and then by t = 0, the returns recover and start going up again (basically, an over-reaction to bad news that has a 5-day lifetime)?

This was what I was hoping to get across. I have a hunch that much of the information we see in news articles are already incorporated into stock prices before the articles get published (asymmetric distribution of information) - Paul Tetlock talks about it here: https://www0.gsb.columbia.edu/faculty/ptetlock/papers/Tetlock%2004%2010%20News%20and%20Asymmetric%20Information.pdf

In other words, how should I interpret t = 0? Naturally, one would think it is the time at which negative news & sentiment were published, but is it?

Pretty much in 95% of all cases you can assume that the asof_date is symmetric to the publish of a news article. I think the returns decline (starting from t=-5) is more a phenomenon of asymmetric news versus a discrepancy between dates in news article publish & sentiment publish.

"Sell the rumor, buy the news"

Thanks Seong,

Interesting. One has to wonder how the "rumor" process works. Do you know anything about the nature of the articles that are published? Are they generally commentary on news releases by the companies themselves ("Acme Corporation announced today that it will no longer sell dynamite to coyotes, as part of its class action suit settlement, which was brought on by a group of road runners.")? Or are the articles independent investigative reporting? My guess is that we are talking about corporate announcements/news releases, primarily, but I could be wrong.

Also, who is acting on the rumors? Are mutual funds/ETFs/pension funds/etc.--big players--shifting allocations based on rumors? Or some other market players? And how are they getting the information?

Seems like there could be some illegal activity here, too. If the "rumor" is knowledge of the news release, and trades are made based on that information, then I ain't no lawyer, but it would seem like trading on material non-public information. It sure seems like we are talking about a form of insider trading here, no?

A suggestion:
How about filtering the articles and blogs based on some category for example the viewership, or number of shares of the article as well as the profile and authenticity of the article/blog so that we may get a better estimate ?

Regards
Asad

Grant,

I'm not sure if rumors are the main reason for the drop, although I could be very wrong.

There are some cases where the asymmetric distribution might not be because of certain investors having access to the information before everyone else but rather because people simply pay less attention to certain stocks. IE. news articles about a less liquid security is going to get less attention, slower absorption rates (literally, it takes people a lot longer to figure out that something happened to FORD - Forward Industries - than F - Ford Motor Company), and such.

However I think trading on material non-public information. It sure seems like we are talking about a form of insider trading here, no?also plays a part.

As for this question: Do you know anything about the nature of the articles that are published? Kumesh talks about the filtering methods around 29:00 of the recorded webinar (https://www.youtube.com/watch?v=GlEV3WtquK4&feature=youtu.be) where essentially, Accern does only act on information that's relevant to a security's stock price.

I would guess the major funds that use a news based strategy use a NLP algorithm and other methods of automated reading. For the purposes of this strategy they probably treat each stock as if it were an ETF. For example, for MSFT if there is news about consumer, USD, EURO, bonds (not sure if still true) and others which they then play a MSFT (market neutral) on any of the news sectors that would affect Microsoft.
They probably also rate the news for strength, novelty, and subjectiveness. At the objectively rated side of the scale they hold maybe ten minutes, and for the very subjective, they hold up to three days.