Textual Analysis of Stock Market Prediction Using Breaking
Financial News: The AZFinText System
Robert P. Schumaker and Hsinchun Chen
Artificial Intelligence Lab, Department of Management Information Systems
The University of Arizona, Tucson, Arizona 85721, USA
Word Count: 7963
Our research examines a predictive machine learning approach for financial news articles
analysis using several different textual representations: Bag of Words, Noun Phrases, and Named
Entities. Through this approach, we investigated 9,211 financial news articles and 10,259,042
stock quotes covering the S&P 500 stocks during a five week period. We applied our analysis to
estimate a discrete stock price twenty minutes after a news article was released. Using a Support
Vector Machine (SVM) derivative specially tailored for discrete numeric prediction and models
containing different stock-specific variables, we show that the model containing both article
terms and stock price at the time of article release had the best performance in closeness to the
actual future stock price (MSE 0.04261), the same direction of price movement as the future
price (57.1% directional accuracy) and the highest return using a simulated trading engine
(2.06% return). We further investigated the different textual representations and found that a
Proper Noun scheme performs better than the de facto standard of Bag of Words in all three
Categories and Subject Descriptors: G.1.10 [Numerical Analysis] Applications; H.4.2
[Information Systems Applications] Decision Support; I.2.7 [Natural Language Processing]
General Terms: Algorithms, Design, Experimentation, Measurement, Performance
Additional Key Words and Phrases: SVM, Stock Market, Prediction
Portions of this article were previously published as Schumaker, R., & Chen, H., (2006) Textual Analysis of Stock
Market Prediction Using Financial News Articles. 12th Americas Conference on Information Systems (AMCIS)
Stock Market prediction has always had a certain appeal for researchers. While
numerous scientific attempts have been made, no method has been discovered to accurately
predict stock price movement. The difficulty of prediction lies in the complexities of modeling
market dynamics. Even with a lack of consistent prediction methods, there have been some mild
Stock Market research encapsulates two elemental trading philosophies; Fundamental
and Technical approaches . In Fundamental analysis, Stock Market price movements are
believed to derive from a security’s relative data. Fundamentalists use numeric information such
as earnings, ratios, and management effectiveness to determine future forecasts. In Technical
analysis, it is believed that market timing is key. Technicians utilize charts and modeling
techniques to identify trends in price and volume. These later individuals rely on historical data
in order to predict future outcomes.
One area of limited success in Stock Market prediction comes from textual data.
Information from quarterly reports or breaking news stories can dramatically affect the share
price of a security. Most existing literature on financial text mining relies on identifying a
predefined set of keywords and machine learning techniques. These methods typically assign
weights to keywords in proportion to the movement of a share price. These types of analyses
have shown a definite, but weak ability to forecast the direction of share prices.
In this paper we experiment using several linguistic textual representations, including
Bag of Words, Noun Phrases, and Named Entities approaches. We believe that combining more
precise textual representations with past stock pricing information will yield improved
This paper is arranged as follows. Section 2 provides an overview of literature
concerning Stock Market prediction, textual representations, and machine learning techniques.
Section 3 describes our research questions. Section 4 outlines our system design. Section 5
provides an overview of our experimental design. Section 6 expresses our experimental findings
and discusses their implications. Section 7 delivers our experimental conclusions with a brief
oratory on future directions for this stream of research.
2 Literature Review
When predicting the future prices of Stock Market securities, there are several theories
available. The first is Efficient Market Hypothesis (EMH) . In EMH, it is assumed that the
price of a security reflects all of the information available and that everyone has some degree of
access to the information. Fama’s theory further breaks EMH into three forms: Weak, Semi-
Strong, and Strong. In Weak EMH, only historical information is embedded in the current price.
The Semi-Strong form goes a step further by incorporating all historical and currently public
information in the price. The Strong form includes historical, public, and private information,
such as insider information, in the share price. From the tenets of EMH, it is believed that the
market reacts instantaneously to any given news and that it is impossible to consistently
outperform the market.
A different perspective on prediction comes from Random Walk Theory . In this
theory, Stock Market prediction is believed to be impossible where prices are determined
randomly and outperforming the market is infeasible. Random Walk Theory has similar
theoretical underpinnings to Semi-Strong EMH where all public information is assumed to be
available to everyone. However, Random Walk Theory declares that even with such
information, future prediction is ineffective.
It is from these theories that two distinct trading philosophies emerged; the
fundamentalists and the technicians. In a fundamentalist trading philosophy, the price of a
security can be determined through the nuts and bolts of financial numbers. These numbers are
derived from the overall economy, the particular industry’s sector, or most typically, from the
company itself. Figures such as inflation, joblessness, return on equity (ROE), debt levels, and
individual Price to Earnings (PE) ratios can all play a part in determining the price of a stock.
In contrast, technical analysis depends on historical and time-series data. These
strategists believe that market timing is critical and opportunities can be found through the
careful averaging of historical price and volume movements and comparing them against current
prices. Technicians also believe that there are certain high/low psychological price barriers such
as support and resistance levels where opportunities may exist. They further reason that price
movements are not totally random, however, technical analysis is considered to be more of an art
form rather than a science and is subject to interpretation.
Both fundamentalists and technicians have developed certain techniques to predict prices
from financial news articles. In one model that tested trading philosophies; LeBaron et. al.
posited that much can be learned from a simulated stock market with simulated traders . In
their work, simulated traders mimicked human trading activity. Because of their artificial nature,
the decisions made by these simulated traders can be dissected to identify key nuggets of
information that would otherwise be difficult to obtain. The simulated traders were programmed
to follow a rule hierarchy when responding to changes in the market; in this case it was the
introduction of relevant news articles and/or numeric data updates. Each simulated trader was
then varied on the timing between the point of receiving the information and reacting to it. The
results were startling and found that the length of reaction time dictated a preference of trading
philosophy. Simulated traders that acted quickly formed technical strategies, while traders that
possessed a longer waiting period formed fundamental strategies . It is believed that the
technicians capitalized on the time lag by acting on information before the rest of the traders,
which lent this research to support a weak ability to forecast the market for a brief period of time.
In similar research on real stock data and financial news articles, Gidofalvi gathered over
5,000 financial news articles concerning 12 stocks, and identified this brief duration of time to be
a period of twenty minutes before and twenty minutes after a financial news article was released
. Within this period of time, Gidofalvi demonstrated that there exists a weak ability to predict
the direction of a security before the market corrects itself to equilibrium. One reason for the
weak ability to forecast is because financial news articles are typically reprinted throughout the
various news wire services. Gidofalvi posits that a stronger predictive ability may exist in
isolating the first release of an article. Using this twenty minute window of opportunity and an
automated textual news parsing system, the possibility exists to capitalize on stock price
movements before human traders can act.
2.1 Textual Representation
There are a variety of methods available to analyze financial news articles. One of the
more common methods is to apply a vector representation where article terms are indexed and
then weighted. Selecting article terms can be as simple as tokenizing and using each word in the
document. This technique assigns importance to determiners and prepositions which have little
contribution to the overall meaning of the article. One method of circumventing these problems
is to use a Bag of Words approach. In this approach, a list of semantically empty stop-words are
removed from the article (e.g.; the, a, and for). The remaining terms are then used as the textual
representation. The Bag of Words approach has been used as the de facto standard of financial
article research primarily because of its simple nature and its ability to produce a suitable
representation of the text.
Building upon the Bag of Words approach, another tactic is to use a subset of terms as
features , which can address issues related to article scaling while still encompassing the
important concepts of an article . One such method using this approach is Noun Phrasing.
Noun Phrasing is accomplished through the use of a syntax where parts of speech (i.e., nouns)
are identified through the aid of a lexicon and aggregated using syntactic rules on the
surrounding parts of speech, forming noun phrases.
A third method of article representation is Named Entities. This technique builds upon
Noun Phrases by using lexical semantic/syntactic tagging where nouns and noun phrases can be
classified under predetermined categories . This contrasts with using a differential approach,
where concepts can be determined using a distributional analysis . An example of the
predetermined category approach is the MUC-7 framework of entity classification, where
categories include date, location, money, organization, percentage, person and time. The entity
tagging procedure can happen in a number of ways. Typically, successful taggers have large
lexicons of sample entities and/or word patterns, which may include both syntax and lexical
information. Lexicons and pattern information can be used to designate features or machine
learning approaches or be incorporated in rule in more of a rules-based approach. The large
quantities of patterns are considered relatively cheap or shallow knowledge to obtain. Thus
reusability of the extraction rules is not a priority. When input text is matched to the stored
extraction patterns the corresponding input text gets assigned an entity tag. Because of the
constrained categories, Named Entities in effect provide the smallest coverage of the document,
but identify very specific types of phrases, which may or may not be helpful for stock price
Both Noun Phrases and Named Entities have shown limited success through previous
comparison trials of tagging accuracy between differing algorithms. However, their usage as
wide-scale textual representations for machine learning purposes remains somewhat unknown.
2.2 Machine Learning Algorithms
Like textual representation, there are also a variety of machine learning algorithms
available. Almost all techniques start off with a technical analysis of historical security data by
selecting a recent period of time and performing linear regression analysis to determine the price
trend of the security. From there, a Bag of Words analysis is used to determine the textual
keywords. Some keywords such as ‘earnings’ or ‘loss’ can lead to predictable outcomes which
are then classified into stock movement prediction classes such as up, down, and unchanged.
Much research has been done to investigate the various techniques that can lead to stock price
classification. Table 1 illustrates a Stock Market prediction landscape of the various machine
Table 1. Prior algorithmic research
From Table 1, several items become readily noticeable. The first of which is that a
variety of algorithms have been used. The second is that almost all instances commonly classify
Algorithm Classification Source Material Examples
Genetic Algorithm 2 categories Undisclosed number of chatroom postings Thomas & Sycara, 2002
3 categories Over 5,000 articles borrowed from Lavrenko Gidofalvi et al. 2001
5 categories 38,469 articles Lavrenko et al. 2000
5 categories 6,239 articles Seo et al. 2002
3 categories About 350,000 articles Fung et al. 2002
3 categories 6,602 articles Mittermayer, 2004
predicted stock movements into a set of classification categories, not a discrete price prediction.
Lastly, not all of the studies were conducted on financial news articles, although a majority was.
The first technique of interest is the Genetic Algorithm. In this study, discussion boards
were used as a source of independently generated financial news . In their approach,
Thomas and Sycara attempted to classify stock prices using the number of postings and number
of words posted about an article on a daily basis. It was found that positive share price
movement was correlated to stocks with more than 10,000 posts. However, discussion board
postings are quite susceptible to bias and noise.
Another machine learning technique, Naïve Bayesian, represents each article as a
weighted vector of keywords . Phrase co-occurrence and price directionality is learned from
the articles which lead to a trained classification system. One such problem with this style of
machine learning is from a company mentioned in passing. An article may focus its attention on
some other event and superficially reference a particular security. These types of problems can
cloud the results of training by unintentionally attaching weight to a casually-mentioned security.
One of the more interesting machine learners is Support Vector Machines (SVM). In the
work of Fung et. al., regression analysis of technical data is used to identify price trends while
SVM analysis of textual news articles is used to perform a binary classification in two predefined
categories; stock price rise and drop . In cases where conflicting SVM classification ensues,
such as both rise and drop classifiers are determined to be positive, the system returns a ‘no
recommendation’ decision. From their research using 350,000 financial news articles and a
simulated Buy-Hold strategy based upon their SVM classifications, they showed that their
technique of SVM classification was mildly profitable.
Mittermayer also used SVM is his research to find an optimal profit trading engine .
While relying on a three tier classification system, this research focused on empirically
establishing trading limits. It was found that profits can be maximized by buying or shorting
stocks and taking profit on them at 1% up movement or 3% down movement. This method
slightly beat random trading by yielding a 0.11% average return.
Many of the prior studies were classification oriented with questions asked such as; will
this article cause the stock price to increase/decrease? These studies were all tests of directional
movement and not the predictors of stock prices. Discrete prediction from numeric trends is
hardly new. However, the application of this regression technique to SVM mechanics is rather
recent . One such method is Sequential Minimal Optimization (SMO) , where many of
the scalability problems from using large training sets has been obviated through a more
simplistic SVM solving technique. This combination of techniques has lead to the completely
numeric prediction studies for futures contracts , but discrete prediction has not been coupled
with a systematic study of various textual analysis methods before.
From prior studies on the textual representation of documents, Joachims posits that
limiting the inclusion of features to three or more instances per document will avoid the problem
of unmanageably large feature spaces . Extending this to textual representation, each feature
is further represented in binary as either a zero or one; the term is either present or not present in
the article . This simple representational scheme is easy to implement and will lead to a
sparse dataset with many zero features.
Applying these regression based methods and textual representation techniques to a
supervised machine learning algorithm such as SVM can lead to a trained system with discrete
Evaluation of output has been generally focused on only one of the following three
metrics; measures of Closeness, Directional Accuracy, or Simulated Trading. In measures of
Closeness, the estimated value from machine learning is compared against the actual value in a
Mean Squared Error (MSE) measure . Directional Accuracy was the more common measure
of previous financial studies, where the direction of the predicted value is compared with the
movement direction of the actual value . Whereas Simulated Trading initiates a simple
trading engine to capitalize on large predicted value differences .
2.3 Financial News Article Sources
In real-world trading applications, the amount of textual data available to stock market
traders is staggering. This data can come in the form of required shareholder reports,
government-mandated forms, or news articles concerning a company’s outlook. Articles and
reports are also routinely cross-posted in many different locations leading to problems of
uniqueness and database selection . Reports of an unexpected nature can lead to wildly
significant changes in the price of a security. Table 2 illustrates some examples of textual
Textual Source Types Examples Description
8K Reports on significant changes
10K Annual reports
Recommendations Buy/Hold/Sell assessments
Stock Alerts Alerts for share prices
Financial Times Financial News stories
Wall Street Journal Financial News stories
PRNewsWire Breaking financial news articles
45 financial news wire sources
The Motley Fool Forum to share stock-related
Table 2. Examples of textual financial data
Textual data can arise from two sources; company generated and independently
generated sources. Company generated sources such as quarterly and annual reports can provide
a rich linguistic structure that if properly read can indicate how the company will perform in the
future . This textual wealth of information may not be explicitly shown by financial ratios
but rather encapsulated in forward-looking statements or other textual locations. Independent
sources such as analyst recommendations, news outlets, and wire services can provide a more
balanced view of the company and have a lesser potential to bias news reports. Discussion
boards can also provide independently generated financial news; however, they can be suspect
News outlets can be differentiated from wire services in several different ways. One of
the main differences is that news outlets are centers that publish available financial information
at specific time intervals. Examples include Bloomberg, Business Wire, CNN Financial News,
Dow Jones, Financial Times, Forbes, Reuters, and the Wall Street Journal [3, 23]. In contrast,
news wire services publish available financial information as soon as it is publicly released or
discovered. News wire examples include PRNewsWire, which has free and subscription levels
for real-time financial news access, and Yahoo Finance, which is a compilation of 45 news wire
services including the Associated Press and PRNewsWire. Besides their relevant and timely
release of financial news articles, news wire articles are also easy to automatically gather and are
an excellent source for computer-based algorithms.
Stock Quotations are also an important source of financial information. Quotes can be
divided into various increments of time from minutes to days, however, one minute increments
provide sufficient granularity for machine learning.
While previous studies have mainly focused on the classification of stock price trends,
none has been discovered to harness machine learning to determine a discrete stock price
prediction based on breaking news articles. Prior techniques have relied solely on a Bag of
Words approach and not other textual representations. Finally, there is no consensus on what
information to include in a model that will lead to better performance. From these gaps in the
research we form the crux of our study with the following questions.
3 Research Questions
Given that prior research in textual financial prediction has focused solely on the
classification of stock price direction, we ask whether the prediction of discrete values is
possible. This leads to our first research question.
• How effective is the prediction of discrete stock price values using textual
financial news articles?
We expect to find that discrete prediction from textual financial news article is possible.
Since prior research has indicated that certain keywords can have a direct impact on the
movement of stock prices, we believe that predicting the magnitude of these movements is
Prior research into stock price classification has almost exclusively relied on a Bag of
Words approach. While this de facto standard has led to promising results, we feel that other
textual representation schemes may provide better predictive ability, leading us to our second
• Which combination of textual analysis techniques is most valuable in stock price
Since prior research has not examined this question before, we are cautious in answering
such an exploratory issue. However, we feel that other textual representation schemes may serve
to better distill the article into its essential components.
4 System Design
From these questions we developed the AZFinText system illustrated in Figure 1.
Figure 1. AZFinText system design
In this design, each financial news article is represented using three textual analysis
techniques; Bag of Words, Noun Phrases, and Named Entities. These representations identify
the important article terms and store them in the database. To limit the size of the feature space,
we selected terms that occurred three or more times in a document .
To perform our textual analysis we chose a modified version of the Arizona Text
Extractor (AzTeK) system which performs semantic/syntactic word level tagging as well as
phrasal aggregation. AzTeK’s Noun Phrasing component works by using a syntactic tagger to
identify and aggregate the document’s noun phrases and was found to have an 85% F-measure
for both precision and recall, which is comparable to other tools . The Entity extractor
portion goes one step further by assigning hybrid semantic/syntactic tags to document terms and
phrases in one of the seven predefined categories of date, location, money, organization,
percentage, person and time . These entities are then identified through the usage of a
lexicon. Although the AzTeK system was selected due to availability, it performs adequately for
noun phrase and named entity extraction. However, there are many other such systems as
reported in the Message Understanding Conference , that can be adopted for financial news
Stock Quotes are gathered on a per minute basis for each stock. When a news article is
released, we estimate what the stock price would be 20 minutes after the article was released. To
do this we perform linear regression on the quotation data using an arbitrary 60 minutes prior to
article release and extrapolate what the stock price should be 20 minutes in the future.
To test the types of information that need to be included, we developed four different
models and varied the data given to them. The first model, Regress, was a simple linear
regression estimate of the +20 minute stock price. Assuming that breaking financial news
articles have no impact on the movement of stock prices, we would expect a reasonable
performance from this model. While we acknowledge the obvious violation of Random Walk
Theory, within such a compressed amount of time weak predictive ability remains . Next the
three models use the supervised learning of SVM regression to compute their +20 minute
predictions. Model M1, uses only extracted article terms for its prediction. While no baseline
stock price exists within this model, we chose it because of its frequent usage in prior studies on
directional classification of stock prices. Model M2, uses extracted article terms and the stock
price at the time the article was released. We feel that given a baseline of stock price that this
model will fare better. Model M3, uses extracted terms and a regressed estimate of the +20
minute stock price. This model may lead to better predictive results should the article terms have
no impact on the movement of the stock price. All three models rely on using article terms in
their prediction. SVM learns what terms lead to share price changes and adjust their weights
depending on the severity of price changes.
To illustrate how the AZFinText system works, we offer a sample news article  and
step through the logic of our system.
Schwab shares fell as much as 5.3 percent in morning trading on the New York Stock Exchange
but later recouped some of the loss. San Francisco-based Schwab expects fourth-quarter profit of
about 14 cents per share two cents below what it reported for the third quarter citing the impact of
fee waivers a new national advertising campaign and severance charges. Analysts polled by
Reuters Estimates on average had forecast profit of 16 cents per share for the fourth quarter. In
September Schwab said it would drop account service fees and order handling charges its seventh
price cut since May 2004. Chris Dodds the company s chief financial officer in a statement said
the fee waivers and ad campaign will reduce fourth-quarter pre-tax profit by $40 million while
severance charges at Schwab s U.S. Trust unit for wealthy clients will cut profit by $10 million.
The NYSE fined Schwab for not adequately protecting clients from investment advisers who
misappropriated assets using such methods as the forging of checks and authorization letters. The
improper activity took place from 1998 through the first quarter of 2003 the NYSE said. This case
is a stern reminder that firms must have adequate procedures to supervise and control transfers of
assets from customer accounts said Susan Merrill the Big Board s enforcement chief. It goes to the
heart of customers expectations that their money is safe. Schwab also agreed to hire an outside
consultant to review policies and procedures for the disbursement of customer assets and detection
of possible misappropriations the NYSE said. Company spokeswoman Alison Wertheim said
neither Schwab nor its employees were involved in the wrongdoing which she said was largely the
fault of one party. She said Schwab has implemented a state-of-the-art surveillance system and
improved its controls to monitor independent investment advisers. According to the NYSE
Schwab serves about 5 000 independent advisers who handle about 1.3 million accounts.
Separately Schwab said October client daily average trades a closely watched indicator of
customer activity rose 10 percent from September to 258 900 though total client assets fell 1
percent to $1.152 trillion. Schwab shares fell 36 cents to $15.64 in morning trading on the Big
Board after earlier falling to $15.16. (Additional reporting by Dan Burns and Karey Wutkowski)
Figure 2. Example AZFinText representation
The first step in our system is to extract the text from each article using our three textual
representations, independently. This builds up 3 separate corpora for each representation. Then
each article is passed through the system, one at a time, as is shown in Figure 2 with the prior
Schwab article using the Bag of Words representation. In the Textual Analysis box, extracted
terms are represented in binary as either present or not in this article. Supposing our corpus also
contained the term Reuters that appeared in many different articles but not this one, the term is
given a zero for not being present in the current article. For stock quotation data, we lookup
what the stock price was at the time of article release ($15.65), calculate a regression estimate of
the +20 minute stock price over the past hour ($15.633) and lookup the actual +20 minute stock
price for training and later evaluation ($15.59). This data is then taken to the Model Building
stage where the various models are given their appropriate data. Following that, machine
learning takes place and an estimate of the +20 minute stock price is produced ($15.645). We
can see from the stock prices given in this example that Schwab’s share price dropped six cents
while the model estimate figures a more conservative half penny drop.
5 Experimental Design
For our experiment we picked a research period of Oct. 26 to Nov. 28, 2005 to gather
news articles and stock quotes. We further focused our attention only on companies listed in the
S&P 500 as of Oct. 3, 2005. We acknowledge that several mergers and acquisitions did take
place during this period of time; however, this only had an effect on less than 2% of the stocks
tracked. In order to eliminate the ‘company in passing’ problem, we gathered the news articles
from Yahoo Finance using a company’s stock ticker symbol. This resulted in articles on 484 of
the 500 companies listed in the S&P 500. Articles were further constrained to a time frame of
one hour after the stock market opened to twenty minutes before the market closed. This period
of time allows for sufficient data to be gathered for prior regression trend analysis and future
estimation purposes. We further limited the influence of articles such that we did not use any
two or more articles that occurred within twenty minutes of each other. This measure eliminated
several possible avenues of confounding results.
By performing these actions we gathered 9,211 candidate financial news articles and
10,259,042 stock quotes over the five-week period. From this pool of news articles we analyzed
them using the three textual representations and retained only those terms that appeared three or
more times in an article, which results in a differing number of articles. The filtering process
resulted in the following breakdown:
• Bag of Words used 4,296 terms from 2,839 articles
• Noun Phrases used 5,283 terms from 2,849 articles
• Named Entities used 2,856 terms from 2,620 articles
Article and stock quote data was then processed by a Support Vector Machine derivative,
using Sequential Minimal Optimization  in a form of regression, which can handle discrete
number analysis .
Following training, we chose three evaluation metrics; Closeness, Directional Accuracy,
and a Simulated Training Engine. The Closeness metric evaluated the difference between the
predicted value and the actual stock price, measured using Mean Squared Error (MSE).
Directional Accuracy measured the up/down direction of the predicted stock price compared
with the actual direction of the stock price. While the inclusion of Directional Accuracy may not
seem intuitive given the measure of Closeness, it is possible to be close in prediction yet predict
the wrong direction of movement. This leads us to a third evaluation measure using a Simulated
Trading Engine that invests $1,000 per trade and follows simple trading rules. The rules
implemented by our trading engine are a modified version of those proposed by Mittermayer to
maximize short-term trading profit . Our Simulated Trading Engine evaluates each news
article and will buy/short the stock if the predicted +20 minute stock price is greater than or
equal to 1% movement from the stock price at the time the article was released. Any
bought/shorted stocks are then sold after 20 minutes. This assumes a zero transaction cost which
is consistent with the research of Lavrenko [12, 13] and Mittermayer  who argue that trading
in volume will offset the costs of trading.
6 Experimental Findings and Discussion
In order to answer our research questions on the effectiveness of discrete stock prediction
and the best textual representation; we tested our three models against a regression-based
predictor using the three dimensions of analysis; measures of Closeness, Directional Accuracy
and a Simulated Trading Engine. Table 3 shows the results of the Closeness measures where
smaller numbers indicate less error in price prediction. Table 4 illustrates Directional Accuracy,
where 50.0% could be achieved by chance alone. Finally, Table 5 displays the returns obtained
from the Simulated Trading Engine.
Bag of Words 0.07279 930.87 0.04422 0.12605
Noun Phrases 0.07279 863.50 0.04887 0.17944
Average 0.07212 848.15 0.04261 0.12893
Table 3. Closeness results
Bag of Words 54.8% 52.4% 57.0% 57.0%
Noun Phrases 54.8% 56.4% 58.0% 56.9%
Totals 54.6% 54.6% 57.1% 56.9%
Table 4. Directional Accuracy results
Bag of Words -1.81% -0.34% 1.59% 0.98%
Noun Phrases -1.81% 0.62% 2.57% 1.17%
Totals -1.95% -0.05% 2.06% 1.67%
Table 5. Simulated Trading Engine results
A. Model M2 with articles terms and baseline stock price performed the best.
From looking at the average results in Table 3, Model M2 which used both article terms
and the stock price at the time of article release, had the lowest MSE score (0.04261) of any of
the models (p-values < 0.05). This result signifies that Model M2’s predictions were closer to
the actual +20 minute stock price than any of the other models including linear regression
(Regress). Looking deeper into the results, we find that Model M2 performed better than
Regress in each of the three textual representations, which supports Gidofalvi’s claim of weak
Model M1, which used only article terms, had a difficult time in its estimation of future
stock prices with an average Closeness score of 848.15. While this model may have been
appropriate for prior classification-only studies, this unpleasant value was somewhat expected
given the lack of baseline stock prices.
The other item of interest is that Model M2’s Named Entities representation had the best
performance at 0.03407 (p-values < 0.05). We will further investigate the effects of textual
representation in a later section.
Turning our attention back to Model M2, we examined the weighting scheme that SVM
assigned to the training variables. The stock price at the time of article release was given a
weight of 0.9997 by SVM, while the article terms had a combined weight of 0.0003. While the
weighting of article terms may appear superficially light, these terms are important because they
provide the final touches to the estimated +20 minute stock price. If we were to rely on stock
price alone without article terms, we would have the values of Regress, and Model M2 which
used both the stock price and article terms performed better than Regress. This signifies that the
0.0003 combined weighting of article terms is an important element in providing more accurate
price results. If instead we used a regressed price estimate plus article terms, we would have
Model M3, and M2 performed better than M3. However, article terms alone were not sufficient
in estimating the future stock price as demonstrated by Model M1.
In order to gain insight into the performance of our results, we can compare them to Pai
and Lin whom conducted a similar study on forecasting stock prices . In their study, they
attempted stock price prediction one day in advance, using a small set of stocks and only close of
day prices. They managed an MSE score of 0.3001 and comparing that to our average MSE
scores in Table 3, our findings were an order of magnitude better.
In evaluating the Directional Accuracy results of Table 4, we again note that Model M2
performed better on average (57.1%) than the other models (p-values < 0.05). Regress did not
perform so well (54.6%), which would seem to indicate that unexpected stock swings were
captured by article terms. Comparing our results to previous studies shows that our values are
somewhat reasonable. Cho et. al. which used 100 days of training articles and 392 keywords,
had an average directional accuracy of 46.8% .
In the Simulated Trading results of Table 5, Model M2 using article terms and the stock
price at the time of article release again had the best performance at 2.06% return (p-values <
0.05). This result would imply that Model M2 was better able to capitalize on trading
opportunities given article terms and baseline stock price. Comparing this model against Model
M1 with a trading return of -0.05%, we see that using article terms alone were insufficient. The
results from Regress (-1.95% return) were unexpected. We believe that this finding was from
the news articles themselves affecting major changes in the share price of stocks. Correlating
our results with prior studies, Lavrenko et. al. claimed a 2% return from tracking four stocks over
a forty day period . In a similar study, Lavrenko et. al. expanded the number of stocks to
127 over the same 40 day period and had a much lower return of 0.23% . Both of these
studies used essentially the same trading mechanism as we did which leads to an interesting
observation; that perhaps more stocks lead to lower returns, although our study tracked 500
companies over 23 trading days. In a third simulated trading study, Mittermayer obtained a
0.11% average return using all of the stocks from Nasdaq, NYSE, and AMEX over a one year
period of time. From our results, it would appear that our system is achieving fairly reasonable
results at a 2.09% return.
Compiling all of the model performances together, Model M2 using article terms and the
stock price at the time of article release performed best in all three metrics; measures of
Closeness (0.04261), Directional Accuracy (57.1%), and Simulated Trading (2.06% return).
This model was better able to capture stock price movements and further bolsters the idea of
weak short-term predictability. Our results were also inline with those from prior studies and
mostly performed better. With some tweaking to how we classify directional movement, we feel
that our system could produce better Directional Accuracy results as well.
B. A superset of Named Entities was the best textual representation.
To answer our second research question, which combination of textual analysis
techniques is most valuable in stock price prediction, we compare the averages of each textual
representation using our three metrics. Table 6 presents the results of Closeness measures, Table
7 displays Directional Accuracy and Table 8 illustrates the Simulated Trading Engine.
Bag of Words 0.07279 930.87 0.04422 0.12605 232.77789
Noun Phrases 0.07279 863.50 0.04887 0.17944 215.95020
Table 6. Closeness results
Bag of Words 54.8% 52.4% 57.0% 57.0% 55.3%
Noun Phrases 54.8% 56.4% 58.0% 56.9% 56.5%
Table 7. Directional Accuracy results
Bag of Words -1.81% -0.34% 1.59% 0.98% 0.10%
Noun Phrases -1.81% 0.62% 2.57% 1.17% 0.64%
Table 8. Simulated Trading Engine results
From these tables, Named Entities had the lowest score in measures of Closeness
(185.50404) and Noun Phrases had the better score in both Directional Accuracy (56.5%) and
Simulated Trading (0.64%), all p-values < 0.05. These seemingly confusing results were not as
clear-cut as our Model selection in the previous section as no one textual representation
dominated the results.
However, it must be noted that these averaged results contain noise from previously
failed models. If we were to focus only on the textual results for Model M2 and discard the other
models, Noun Phrases performed the best in 2 of the 3 metrics and Named Entities in the
These results ran contrary to our expectations. We had assumed that a Named Entity
representation would generate better performance because of its ability to abstract the article
terms and discard the noise of terms picked up by both Bag of Words and Noun Phrases. This
MUC-7 textual representation was not sufficient to adequately model our article terms and lead
us to ask the question, what were the differences between Noun Phrases and Named Entities?
The answer was that Named Entities are essentially specialized Proper Nouns. The AzTeK
system we used for part of speech tagging, identifies select terms in one of seven categories;
date, location, money, organization, percentage, person and time . Words in these categories
are basically a subset of Noun Phrases. We believe that expanding the number of categories for
Named Entities will lead to a better representational scheme.
In order to investigate this we took a subset of terms from Noun Phrases that were tagged
as Proper Nouns and introduced a fourth, hybrid, textual representation of Proper Nouns. This
selection of terms is a comparable superset of Named Entities but without the Entity categories.
Proper Nouns captured 3,710 article terms from 2,809 articles compared to the 5,283 terms in
2,849 articles for Noun Phrases and 2,856 terms in 2,620 articles for Named Entities.
To give the reader an understanding of what types of terms would be captured as Proper
Nouns and not Named Entities, we refer back to the sample news article immediately preceding
Figure 2. It is important to remember that Named Entities are derived using a semantic lexicon
of previous input. Therefore, terms such as NYSE, which do not appear as a Named Entity, will
be depicted in a Proper Noun representation.
Restating the metrics in terms of Model M2 to clear up some of the noise from the other
failed models, we introduce the following data. Table 9 shows measures of Closeness, Table 10,
Directional Accuracy and Table 11, Simulated Trading.
Bag of Words 0.04422
Noun Phrases 0.04887
Proper Nouns 0.04433
Table 9. Closeness results
Bag of Words 57.0%
Noun Phrases 58.0%
Proper Nouns 58.2%
Table 10. Directional Accuracy results
Bag of Words 1.59%
Noun Phrases 2.57%
Proper Nouns 2.84%
Table 11. Simulated Trading Engine results
The first item of interest is that the Proper Nouns subset performed better than Noun
Phrases in all three metrics; 0.04433 to 0.04887 in measures of Closeness, 58.2% to 58.0% in
Directional Accuracy and 2.84% to 2.57% in Simulated Trading (all p-values < 0.05). This
would seem to back up our initial expectation that a more abstract textual representation would
perform better. In comparison to Named Entities, Proper Nouns performed better in 2 of the 3
metrics, Directional Accuracy and Simulated Trading whereas Named Entities had better success
at measures of Closeness. This would indicate that the direction we have undertaken is perhaps
correct, but is still in need of refinement. We would suggest that future research should evaluate
expanding the number of entity categories and evaluating the optimal mix for business-related
Overall, Bag of Words performed poorly by comparison. While this textual
representation may be the de facto standard used in other studies, its weak performance is
believed to arise from its reliance on too many noisy article terms. Noun Phrases performed
much better with good performance in both Directional Accuracy and Simulated Trading.
However, it suffered from poor Closeness measures. We believe that this is the result of using a
better tuned representational scheme of news articles, as compared to the Bag of Words
approach. Yet Noun Phrases still possessed some elements of noise which led to less than
desirable Closeness scores. Named Entities had some problems and did not perform as expected.
While this representation had the best Closeness score in prediction accuracy, it was unable to
translate those gains into both Directional Accuracy and Simulated Trading returns. This is
probably the result of using a limited set of Entity categories which was unable to fully represent
the content of financial news articles. Finally, Proper Nouns had the better performance results.
While this textual representation can be thought of as the hybrid go-between for Noun Phrases
and Named Entities, it had a solid performance on both Directional Accuracy and Simulated
Trading. This result is likely attributable to Proper Nouns adequately using the article terms in a
manner that was freer of the noise plaguing Noun Phrases and free of the constraining categories
used by Named Entities.
7 Conclusions and Future Directions
Our first conclusion was that Model M2, using both article terms and the stock price at
the time of article release, had a dominating performance in all three metrics; measures of
Closeness at 0.04261, Directional Accuracy at 57.1% and Simulated Trading at a 2.06% return.
These results were the direct consequence of this model’s ability to capitalize on the article terms
and stock price for machine learning.
Our second conclusion was that Proper Nouns had the better textual representation
performance. While it performed best in 2 of the 3 metrics, Directional Accuracy at 58.2% and
Simulated Trading at 2.84%, it pulled up short on measures of Closeness, 0.04433, as compared
to Named Entities with 0.03407, all p-values < 0.05. However, this subset representation
performed better than its parent, Noun Phrases, in all three metrics. We believe that Proper
Nouns can attribute its success to being freer of the term noise used by Noun Phrases and free of
the constraining categories used by Named Entities. Although more research into what
constitutes an optimum mix of entity categories is encouraged.
Future research includes using other machine learning techniques such as Relevance
Vector Regression, which promises to have better accuracy and fewer vectors in classification
. It would also be worthwhile to pursue expanding the selection of stocks outside of the S&P
500. While the S&P 500 is a fairly stable set of companies, perhaps more volatile and less
tracked companies may provide interesting results. Another worthwhile approach would be to
test a model based on article terms and percentage of stock price change. While our models
relied on fixed stock prices that traded within a consistent range, penny stocks with wild
fluctuations may prove worthy of further research. Lastly, while we trained our system on the
entire S&P 500, it would be a good idea to try more selective article training such as industry
groups or company peer group training and examine those results in terms of prediction
Finally, there are some caveats to impart to readers. While the findings presented here
are certainly interesting, we acknowledge that they rely on a small dataset. Using a larger dataset
would help offset any market biases that are associated with using a compressed period of time,
such as the effects of cyclic stocks, earnings reports, mergers and other unexpected surprises.
 Bishop, C.M. and M.E. Tipping, Bayesian Regression and Classification. Advances in
Theory: Methods, Models and Applications, ed. G.H. J. Suykens, S. Basu, C. Micchelli,
and J. Vandewalle. Vol. 190 of NATO Science Series III: Computer & Systems Sciences.
2003, Amsterdam: IOS Press.
 Burns, D. and K. Wutkowski, Schwab to miss forecast, fined by NYSE, in Excerpted
from: http://biz.yahoo.com/rb/051115/financial_schwab.html?.v=3. Nov. 15, 2005:
Retrieved from Yahoo! News.
 Cho, V., Knowledge Discovery from Distributed and Textual Data, in Computer Science.
1999, The Hong Kong University of Science and Technology: Hong Kong.
 Cho, V., B. Wuthrich, and J. Zhang, Text Processing for Classification. Journal of
Computational Intelligence in Finance, 1998. 26.
 Conrad, J.G. and J.R.S. Claussen, Early user - system interaction for database selection in
massive domain-specific online environments. ACM Transactions on Information
Systems, 2003. 21(1): 94-131.
 Fama, E., The Behavior of Stock Market Prices, in Graduate School of Business. 1964,
University of Chicago.
 Fung, G.P.C., J.X. Yu, X. Yu, and W. Lam. News Sensitive Stock Trend Prediction.
Pacific-Asia Conference on Knowledge Discovery and Data Mining (PAKDD). 2002.
 Gao, J.B., S.R. Gunn, C.J. Harris, and M. Brown, A Probabilistic Framework for SVM
Regression and Error Bar Estimation. Machine Learning, 2002. 46(1 - 3): 71-89.
 Gidofalvi, G., Using News Articles to Predict Stock Price Movements. 2001, University
of California, San Diego: Department of Computer Science and Engineering.
 Joachims, T., Text Categorization with Support Vector Machines: Learning with Many
Relevant Features, in Proceedings of the 10th European Conference on Machine
Learning. 1998, Springer-Verlag. 137-142.
 Kloptchenko, A., T. Eklund, J. Karlsson, B. Back, H. Vanharanta, and A. Visa,
Combining Data and Text Mining Techniques for Analysing Financial Reports.
Intelligent Systems in Accounting, Finance & Management, 2004. 12(1): 29-41.
 Lavrenko, V., M. Schmill, D. Lawrie, and P. Ogilvie. Mining of Concurrent Text and
Time Series. The Sixth ACM SIGKDD International Knowledge Discovery and Data
Mining (KDD). 2000b. Boston, MA.
 Lavrenko, V., M. Schmill, D. Lawrie, P. Ogilvie, D. Jensen, and J. Allan. Language
Models for Financial News Recommendation. Proceedings of the 9th International
Conference on Information and Knowledge Management. 2000a.
 Le Moigno, S., J. Charlet, D. Bourigualt, P. Degoulet, and M.-C. Jaulent, Terminology
Extraction from Text to Build an Ontology in Surgical Intensive Care, in AMIA
Symposium. 2002: San Antonio, TX.
 LeBaron, B., W.B. Arthur, and R. Palmer, Time Series Properties of an Artificial Stock
Market. Journal of Economic Dynamics and Control, 1999. 23(9-10): 1487-1516.
 Malkiel, B.G., A Random Walk Down Wall Street. 1973, New York: W.W. Norton &
 McDonald, D.M., H. Chen, and R.P. Schumaker. Transforming Open-Source Documents
to Terror Networks: The Arizona TerrorNet. American Association for Artificial
Intelligence Conference Spring Symposia. 2005. Stanford, CA.
 Mittermayer, M.-A. Forecasting Intraday Stock Price Trends with Text Mining
Techniques. Proceedings of the 37th Hawaii International Conference on Social Systems.
 Moldovan, D., M. Pasca, S. Harabagiu, and M. Surdeanu, Performance issues and error
analysis in an open-domain question answering system. ACM Transactions on
Information Systems, 2003. 21(2): 133-154.
 Pai, P.-F. and C.-S. Lin, A hybrid ARIMA and support vector machines model in stock
price forecasting. Omega, 2005. 33(6): 497-505.
 Platt, J.C., Fast training of support vector machines using sequential minimal
optimization, in Advances in kernel methods: support vector learning. 1999, MIT Press.
 Sekine, S. and C. Nobata. Definition, dictionaries and tagger for Extended Named Entity
Hierarchy. Proceedings of the LREC. 2003.
 Seo, Y.-W., J. Giampapa, and K. Sycara, Text Classification for Intelligent Portfolio
Management. 2002, Carnegie Mellon University: Robotics Institute.
 Tay, F. and L. Cao, Application of Support Vector Machines in Financial Time Series
Forecasting. Omega, 2001. 29: 309-317.
 Technical-Analysis, The Trader's Glossary of Technical Terms and Topics. 2005,
 Thomas, J.D. and K. Sycara. Integrating Genetic Algorithms and Text Learning for
Financial Prediction. Genetic and Evolutionary Computation Conference (GECCO).
2002. Las Vegas, NV.
 Tolle, K.M. and H. Chen, Comparing Noun Phrasing Techniques for Use with Medical
Digital Library Tools. Journal of the American Society for Information Science, 2000.
 Vanschoenwinkel, B., A discrete Kernel Approach to Support Vector Machine Learning
in Language Independent Named Entity Recognition. 2003,
http://citeseer.ist.psu.edu/682269.html: Computational Modeling Lab, Vrije Universiteit,