Retrieving Financial
News Articles
CS 498cxz Text Info Systems
April 27th 2006
Jai Vasanth
Andra Ivan
What is the Goal?
Retrieve news articles that may have a
high impact on a company’s stock price.
– Data acquisition
– Method for assessing an article’s
importance
– Quantifying retrieval performance
Too much Data can be Bad?


Financial news journals:
Cnet News, Google Business News, MSN Money,
Fortune Magazine…
No existing retrieval system for high impact financial
news
– EAnalyst
• Predicts the stock trend for a specific company by associating
news and stock information
• Results are not convincing.

Relevant news articles must exhibit distinctive
language trends
Language Models for Relevance


News articles are treated as a “bag of words”
Relevance scoring schemes:
– Robertson Sparck Jones Model
• Raw tf weighting
Wi  TFi log
• BM25/Okapi approximation
Wi 
pi (1  qi )
qi (1  pi )
TFi (k1  1)
p (1  qi )
log i
k1  TFi
qi (1  pi )
– KL Divergence
Wi  p ( wi | d )  log
– Logistic Regression
– Entropy Maximization
p ( wi|d )
p ( wi | RC )
Maximum Entropy Model




Define n events (features) e1, e2 …en over our
sample space (event ei would be the occurrence of
the ith term)
Let p be the probability distribution that we are trying
to learn.
Maximum Entropy Modeling learns p such that p(e1),
p(e2) ….p(en) are all exactly equal to p(e1), p(e2)
….p(en) where p is the value estimated from the train
set.
The rest of the events are modeled uniformly
(thereby maximizing entropy)
News and Stock Data

Data Collection
– 45,000 news articles published 2000 – 2005 collected (CNet
News).
– Daily stock prices for NASDAQ companies over 2000 – 2005
(Yahoo! Finance)
– NASDAQ index values over 2000 – 2005 (Yahoo! Finance)

Data Preprocessing
– Extract story content from html
– Extract named entities for each news article and match them
against stock symbol (OpenNLP toolkit).
– Tag each article as relevant or not.
– 2000 and 2001 are test sets.
Measuring News Article Impact on Stock
+ 3.03%
5.123 –
2.03 =
3.093 %
NASDAQ: + 2.03
Just this week, Oracle CEO Larry Ellison told the
Financial Times that he would "like to have a
complete stack." Oracle makes billions of dollars
selling databases and business applications. In recent
years, the company has bought up many other
companies, including rivals like PeopleSoft and
Siebel Systems.
+ 5.123%
- 2.04%
RSJ, KL-Divergence
RSJ, KL
relative to NASDAQ index
Logistic Regression, Maximum Entropy
relative to NASDAQ index
Remarks
Imperfect named entity recognition.
 Unigram model is limited in capturing
any semantic meaning of words.
 Naïve approach, not following stories
but word usages.
 Limited data sources.

Top Relevant Articles (RSJ Raw TF)
New top-end Sun chip arriving by October. Unable to meet an earlier
deadline, the company now says the 900MHz UltraSparc III has
passed tests and will ship in workstations in the next 90 days.
Earnings season hits Wall Street this week, with hundreds of
companies giving investors a glimpse of their business results from
the third quarter … Of the 1,262 companies that issued earnings preannouncements this quarter, 782 companies, or 62 percent, warned
of lower earnings ..
Intel's P6 chip architecture not dead yet. Banias, a low-power chip for
notebooks and Net devices, will be based on the older Pentium
processor design that's being phased out in other Intel product lines.
Amazon.com, Compaq Computer, Lucent and a slew of
telecommunications companies are slated to report their quarterly
results. With the exception of Amazon, which said it would top
previous projections, all major companies reporting this week have
lowered the bar for themselves.
Least Relevant Articles
(RSJ Raw TF)
Last week's terrorist attacks marked a significant turning point in the
debate over computer and Internet privacy, giving new weight to
calls for broader government surveillance powers.
Software worms have become the weapons of choice for vandals to
spread their latest creations, surpassing all malicious-code
predecessors in popularity.
Proposed anti-spam measures before Congress this year are facing
opposition from all sides, despite longstanding angst over junk email and high hopes that 2001 would be the year a federal law would
pass. .. said Gary Hermansen, CEO of Brightmail, which provides
spam-filtering technology to Microsoft, EarthLink, AT&T.
And during the past year, several companies including Microsoft,
IBM, 24/7 Media, EarthLink, Excite@Home and DoubleClick have
appointed chief privacy officers, apparently to articulate and enforce
privacy policies
Top Relevant Articles (Logistic Regression)
Lucent Technologies has given Wall Street indications that there may be light
at the end of the tunnel for the financially besieged company, but some
analysts wonder if the worst is yet to come.After a year marked by a series of
financial gaffes, Lucent executives have spent the past few months performing
a massive makeover of a company that was once the darling of Wall Street investors.
The PC industry will begin its quarterly health check later today, when Gateway
reports its second-quarter earnings.The direct PC seller is expected to post earnings
of 36 cents per share and revenue around $2.1 billion when it releases its results
after the close of regular trading. For the same quarter a year ago.
Lucent Technologies' financial woes continued Tuesday as the
telecommunications equipment maker said it expects next quarter's profits will
be lower than expected. In after-hours trading, at 4:40 p.m. PT, Lucent's stock
plummeted $7.25, from $31.38 to $24.13, according to the Island ECN Web
site.
Microsoft at last introduced Windows 2000, an ambitious attempt to make up
ground in the high-end computing world.The business-use software, which the
giant hopes will make inroads on the server computers that power Web sites
while staving off fast-growing Linux, marks an industry milestone because
numerous other companies are expecting to benefit.
Top Nonrelevant Articles (Logistic Regression)
The truth is out there. Just don't go looking for much of it surrounding "Majestic,"
the new online adventure game launched recently by video and computer games
giant Electronic Arts.The game, which draws obvious inspiration from "The XFiles" and the 1997 thriller "The Game," guides players through an interactive
mystery based on actual conspiracy theories involving alien contacts and
government cover-ups.
Venture capitalists recovering from dot-com scare.Though venture capitalists say
their business has undergone a significant chill, they agree that the drop in
temperature is unlikely to threaten--and may even improve--the overall health of
the industry.Adam Zong could almost taste the money.He had just wrapped up a
presentation to a potential investor and was told that a check for $1.5 million
enough to propel his young company
Battle lines harden over Net copyright. A former commissioner of the U.S. Patent
and Trademark Office is stepping back into the digital fray, creating a pro-copyright
coalition aimed at defending his and Congress' work.Bruce Lehman thinks the
digital copyright laws he helped write are in trouble, and it's largely the Net's
fault.Lehman helped author the laws that govern music, video and other digital
media distribution when he ran the U.S. Patent and Trademark Office in the mid1990s.
Virtual shareholder meetings flop. A certain law lets companies hold shareholder
meetings solely online, but critics complain it coshareholders' rights in
jeopardy.You can research stocks, review them, and even purchase them
completely online. But once you've actually bought the stock, you'reback to the
real world.Annual reports, proxies and other corporate documentation are still
shipped out by paper mail to shareholders every year.
Top Relevant Articles (Maximum Entropy)
Earnings march or funeral wake? Earnings season kicks off with a vengeance this
week, as hundreds of companies will tell Wall Street how their businesses fared in
the second quarter.Following are some of the notable tech-related events
scheduled for the week of July 16 through July 20.Earnings season kicks off with
a vengeance this week, as hundreds of companies will tell Wall Street how their
businesses fared in the second quarter.
Rambus at the root of Intel's memory troubles | Intel discovers problems with a
chip, called the memory translation hub, inside several Pentium III-based PCs--a
flaw related to Rambus.Intel's got another problem, and once again the trail leads
to Rambus.Intel today said it has discovered problems with a chip called the
memory translation hub, or MTH, inside several Pentium III-based PCs.
The limbo game will continue next week as investors and analysts see whether
another round of technology bellwethers can go even lower than their
preannouncements have predicted. Amazon.com, Compaq Computer, Lucent and
a slew of telecommunications companies are slated to report their quarterly
results.With the exception of Amazon, which said it would top previous
projections, all major companies reporting this week have lowered the bar for
themselves.
Earnings season kicks into high gear with bellwethers such as IBM, Intel, AMD,
Gateway and Siebel Systems all scheduled to report their first-quarter
numbers.With the notable exception of IBM, most of the big-name chip and PC
makers have already watered down their sales and earnings expectations for the
quarter after a dramatic slowdown in technology spending by companies large
References

Auton Lab Toolkit for Logistic Regression.[www.autonlab.org]

Lavrenko, Mining of Concurrent Text and Time Series.
[http://ciir.cs.umass.edu/~lavrenko/aenalyst/pitch.pdf]

Lucene Search Engine Library.
[http://lucene.apache.org/java/docs/index.html]

Maximum Entropy Modeling Toolkit.
[http://homepages.info.ed.ac.uk/s0450736/maxent_toolkit.html]


Berger, Adam et al. Maximum Entropy Approach to Natural
Language Processing.
Zhai, Cheng. Probabilistic Retrieval Model.
[http://sifaka.cs.uiuc.edu/course/498cxz06s/ppt/probmodel.ppt]