Electronic copy available at: http://ssrn.com/abstract=2580050

News-Based Trading Strategies

Stefan Feuerriegela,, Helmut Prendingerb

aChair for Information Systems Research, University of Freiburg, Platz der Alten Synagoge,79098 Freiburg, Germany

bNational Institute of Informatics (NII), 2-1-2 Hitotsubashi, Chiyoda-ku, Tokyo 101-8430, Japan

Abstract

The marvel of markets lies in the fact that dispersed information is instanta-neously processed by adjusting the price of goods, services and assets. Finan-cial markets are particularly efficient when it comes to processing information;such information is typically embedded in textual news that is then interpretedby investors. Quite recently, researchers have started to automatically deter-mine news sentiment in order to explain stock price movements. Interestingly,this so-called news sentiment works fairly well in explaining stock returns. Inthis paper, we attempt to design trading strategies that are built on textualnews in order to obtain higher profits than benchmark strategies achieve. Es-sentially, we succeed by showing evidence that a news-based trading strategyindeed outperforms our benchmarks by a 9.06-fold performance.

Keywords:Decision support, Financial news, Trading strategies, Text mining, Sentimentanalysis, Trading simulation

1. Introduction

Market efficiency relies, to a large extent, upon the availability of infor-mation. Nowadays, market information can be accessed easily as it comesnavely with the prevalence of electronic markets and, because of the straight-forward access, decision makers can use such information (e. g. [1]) to make

Corresponding author. Mail: stefan.feuerriegel@is.uni-freiburg.de; Tel:+49 761 203 2395; Fax: +49 761 203 2416.

Email addresses: stefan.feuerriegel@is.uni-freiburg.de (Stefan Feuerriegel),helmut@nii.ac.jp (Helmut Prendinger)

Preprint submitted to SSRN March 18, 2015

Electronic copy available at: http://ssrn.com/abstract=2580050

purchases and sales more beneficial. In the same context, the so-called (weak)efficient market hypothesis [2] asserts that financial markets are information-ally efficient, in the sense that financial stock prices accurately reflect all publicinformation at all times; price adjustments occur when previously unknowninformation enters the market, as is the case with news.

Several publications (e. g. [3, 4, 5]) study the market reception of newsannouncements, finding a causal and clearly measurable relationship betweenfinancial disclosures and stock market reaction. Market reception is not onlytriggered by the quantitative facts embedded in financial disclosures, but, moreimportantly, qualitative information drives stock market reactions to financialdisclosures, since news is typically embodied in text messages. In order toextract tone from a textual content, one frequently measures the polarity ofnews by measuring the so-called news sentiment.

While previous research [6, 7, 8, 9] succeeded in establishing a link be-tween news tone and stock market prices, it is not clear how the extracted sen-timent signals can then be utilized to facilitate investment decisions. To closethis gap, this paper studies how news sentiment, as an emergent trend of IT inthe finance industry, can enrich news-based trading. News trading combinesreal-time market data and natural language processing to detect suitable newsannouncements in order to trigger transactions. Its mechanisms are often partof an algorithmic trading system, while many regard it as an enabling DecisionSupport System (DSS) for use in banking and financial markets [10]. Such asystem is depicted in Fig. 1. The systems internal process contains two steps:first, the system extracts the news sentiment as a measure of textual tone. Thisprovides an indicator of the expected return direction. Then, the second stepinvolves executing a trading strategy in order to emit a trading decision.

Sentiment AnalysisSection 3.2

Trading StrategySection 4

Price Momentum

News Announcement

Dictionary

Buy

Short

HoldExpected Return

DirectionDecision

Figure 1: Decision Support System (DSS) transforms news content into trading decisions.

2

With the increasing statistical reliability of text mining algorithms, newsvendors now actively integrate this technology into their traditional platforms.For example, Thomson Reuters offers additional information along with theirnews products, such as Thomson Reuters News Analytics (TRNA) scores, whichmeasure the polarity and novelty of news content. The reasons for this devel-opment predominantly result [10] from recent advances in natural languageprocessing, coupled with cheaper computational power and more diversifiednews sources. Consequently, users are motivated to harness such a DecisionSupport System due to its direct gains, based on the relative value added oradvantage of such a system compared to existing approaches [10].

Consequently, this paper investigates how a Decision Support System canutilize news sentiment to perform stock trading in practice. Several papers [11,12, 13] from the DSS domain elaborate on a general system design in general,but do not go as far as comparing generating trading signals and evaluatingthe accuracy of the triggered decisions within a financial portfolio. Overall,our contribution is as follows: first, we propose different rule-based tradingstrategies. Second, we find quantitative evidence that our news trading sys-tem can outperform benchmark scenarios. In addition, news-based tradingprofits from incorporating other external variables, such as price momentum,to achieve better estimates of possible market reaction in that specific economiccycle. Furthermore, we identify differences in trading performance across firmsectors. More precisely, certain sectors, such as automobile or chemicals, seemto react more to novel announcements, show an improved reliability in theexpected return direction and thus yield higher financial returns.

The remainder of this paper is structured as follows. In Section 2, we reviewrelated research on the sentiment analysis of financial disclosures and newstrading, in which we focus particularly on how both can be tied within a tradingsystem. Next, Section 3 describes the data sources, as well as the news corpus,that is integrated into the sentiment analysis to extract the subjective toneof financial disclosures. The calculated sentiment values are then integrated(Section 4) into various news trading strategies and, finally, Section 5 evaluatesthese strategies in terms of their financial performance.

2. Related Work

In this section, we present related literature grouped into two categories:first, we compare algorithmic approaches that measure news sentiment in fi-nancial disclosures. Second, we review previous works from both IT and fi-

3

nance that distill news into trading actions as part of a decision support systemfor investments.

2.1. News Sentiment in the Financial DomainMethods that use the textual representation of documents to measure the

positivity and negativity of the content are referred to as opinion mining orsentiment analysis. Sentiment analysis can be utilized to extract subjective in-formation from text sources, as well as to measure how market participantsperceive and react to financial materials. In this case, one uses the observedprice reactions following financial text to validate the accuracy of the newssentiment. Based upon sentiment measures, one can study the relationshipbetween financial documents and their effect on markets. Empirical evidence,for example, shows that a discernible relationship between news content andstock market reaction exists, see e. g. [6, 7] for some of the first analyses.

As sentiment analysis is applied to a broad variety of domains and textualsources, research has devised various approaches to measure sentiment. A re-cent literature overview [14] provides a comprehensive domain-independentsurvey and, within the domain of finance, a number of surveys [15, 16] com-pare studies aimed at stock market prediction. For example, dictionary-basedapproaches are very frequently used in recent financial text mining research [17,9, 18, 19, 8]. These methods count the frequency of pre-defined positive andnegative words from a given dictionary, producing results that are straightfor-ward and reliable. In comparison, machine learning approaches [6, 20, 21, 11,for example] offer a broad range of methods, but may suffer from overfit-ting [22].

In this paper, we want to address only the trading simulation itself andso utilize a dictionary-based approach to allow for easier verification of ourresults. Furthermore, dictionary approaches seem to be the more widespreadtechnique nowadays in finance literature.

2.2. News TradingThis section provides a brief overview of components that are necessary for

a news trading system. A more detailed review and taxonomy can be foundin [10]. In behavioral finance, news trading has long been associated with bothnoise and arbitrage trading [23]. Noise traders chase profit from overreactionsto momentary events, whereas arbitrageurs exploit mispricing, possibly indi-cated by news stories [24, 25]. As a consequence, both a lack of informationand misunderstanding may contribute to gains [26, 27] from news trading.

4

Most likely, an even larger advantage would result from automated transac-tions that a Decision Support System could trigger, before human informationprocessing triggered a buy/sell decision, following news releases.

As a first key ingredient for a realistic study of news trading, we need toaccount for incurred transaction fees. Several previous research papers performtrading simulations, but many neglect the importance of transaction fees.

For example, a support vector machine using financial news yields an ac-curacy of 71 % in predicting the direction of asset returns [11]. Overall,this gives an excess return of 2.88 % compared to the S&P 500 index be-tween October 25, 2005 and November 28, 2005. A different approach usesdecision rules with risk words [28] which yields an average annual excessof 20 % compared to U. S. Treasury bills. Similarly, Tetlock [7] incorporatespessimistic words and yields a 7.3 % plus, when compared to the Dow Jones,with the applied trading strategy. However, all the aforementioned papersshare a lack of consideration for transaction fees, which, in fact, can be sub-stantial [29].

One of the few papers that considers transaction fees also utilizes Germanad hoc announcements and yields an accuracy of 65 % when predicting thedirection of returns. Along with transaction fees of 0.1 %, the average re-turn per transaction accounts for 1.1 %. However, this paper [29] relies onone basic strategy (similar to our simple news-based strategy) and does notcompare other trading strategies.

According to [30], trading strategies based on positive and negative sen-timent can be profitable if the transaction costs are moderate. The authorassumes an investor that trades in 62 stocks simultaneously based on Reuterscompany news. In addition, the author considers transaction costs and pos-sible losses from interest rates when no endowment is considered. The find-ings reveal that trades on sell-signals are more profitable but less frequent.Sensitivity is studied by varying transaction costs and risk-free interest ratesbut the author does not evaluate any trading strategies, excepting the navestrategy.

As a second addition to our Decision Support System, we need to test ourtrading strategies using benchmark scenarios. A common approach is to utilizea benchmark stock index for comparison. The author of [31] predicts the di-rection of stock price movements via sparse matrix factorization utilizing news

5

articles from the Wall Street Journal. The results show an accuracy rate of55.7 %, higher than when compared to a reference index. As an alternative,related research also integrates simple buy-and-hold strategies of stocks withthe highest historic returns.

The third component of a news trading system involves trading strategies.Simple buy-and-hold strategies are common when testing news trading in his-toric portfolio simulation. For example, the authors of [32] hypothesize that asentiment-based selection strategy outperforms a classical buy-and-hold bench-mark strategy (holding all stocks over the whole test period). This approachis what the authors call a portfolio selection test. In a different paper [33],the authors build a news categorization and trading system to predict stockprice trends. The system is combined with a trading engine that generatestrading recommendations in the form of buy stock X and hold it until the stockprices hit the +d % barrier. In a similar fashion, a trading strategy can be builtaround the Google query volume [34] for search terms related to finance. Thisvariable is integrated into a simple buy-and-hold strategy (without transactioncosts) to buy the Dow Jones index at the beginning and sell it at the end of thehold period, in which the authors tested various lengths of holding strategies.This strategy yields a 16 % profit, equal to the overall increase in value of theDow Jones index in the time period from January 2004 until February 2011.

All in all, the above references provide evidence that studying trading strate-gies, with a particular focus on news content, is both an intriguing and relevantresearch question for the finance industry. According to [10], it is no surprisethat news trading remains highly specialized and continues to be based on dis-parate decision-making tools and analytical models. Thus, this paper aims toshed light on this research area by comparing different trading strategies interms of their financial performance.

3. Background

This section introduces background knowledge for both datasets and thesentiment analysis. First, we describe the construction of the news corpusthat is used throughout this paper. We then transform this running text intomachine-readable tokens to measure news sentiment.

6

3.1. Data SourcesOur news corpus originates from regulated ad hoc announcements1. These

announcements must be published for all listed firms in Germany in English.We choose this data source primarily because companies are bound to disclosethese ad hoc announcements as soon as possible through standardized chan-nels, thereby enabling us to study the short-term effect of news disclosures onstock prices. In research, ad hoc announcements are a frequent choice [15]when it comes to evaluating and comparing methods for sentiment analy-sis. In addition, this type of news corpus shows several advantages: ad hocannouncements must be authorized by company executives, the content isquality-checked by the Federal Financial Supervisory Authority2 and severalpublications analyze their relevance to the stock market finding a direct re-lationship (e. g. [35, 36, 37]).

Our collected announcements date from the beginning of January 2004until the end of June 2011. We investigate such a long time period to avoidthe possibility of only analyzing news driven predominantly by a single marketevent, for example, the financial crisis. In addition, we apply the following fil-ter rules. First, each announcement must have at least 50 words. Second, wefocus only on ad hoc press releases from German companies which are writtenin the English language. Our final corpus consists of 14,463 ad hoc announce-ments. To study stock market reaction, we use the daily stock market returnsof the corresponding company, originating from Thomson Reuters Datastream.We only include business days and, due to data availability, we yield a total of2108 observations. In addition, we adjust the publication days of ad hoc an-nouncements according to the opening times of the stock exchange. This isachieved by counting all disclosures after 8 p. m. to the next day.

We later also integrate a stock market index into our analysis as follows:in our analysis, the so-called CDAX index works as a benchmark which thetrading strategies need to combat in terms of performance. The CDAX (seeFig. 2) is a German stock market index calculated by Deutsche Brse. It is acomposite index of all stocks traded on the Frankfurt Stock Exchange that arelisted in the General Standard or Prime Standard market segments, giving atotal of 485 stocks. It does not contain foreign firms or foreign stocks, thusserving as a suitable match to our ad hoc news corpus.

1Kindly provided by Deutsche Gesellschaft fr Adhoc Publizitt (DGAP).2Bundesanstalt fr Finanzdienstleistungsaufsicht (BaFin).

7

200

300

400

500

2004 2006 2008 2010

CDAX

Stoc

k Ind

ex (in

Point

s)

Figure 2: The CDAX as a German stock market index representing our benchmark.

3.2. Sentiment AnalysisMethods that use the textual representation of documents to measure posi-

tive and negative content are referred to as opinion mining or sentiment analy-sis [14]. In fact, sentiment analysis can be utilized [15, 16] to extract subjectiveinformation from text sources, as well as to measure how market participantsperceive and react to news. One uses the observed stock price reactions follow-ing a news announcement to validate the accuracy of the sentiment analysisroutines. Thus, sentiment analysis provides an effective tool chain to study therelationship between news content and its market reception [6, 38, 7].

Before performing the actual sentiment analysis, there are several prepro-cessing steps as follows:

1. Tokenization. Corpus entries are split into single words named tokens [39].

2. Negations. Negations invert the meaning of words and sentences [40, 41].When encountering the word no, each of the subsequent three words (i. e.the object) are counted as words from the opposite dictionary. When othernegating terms are encountered (rather, hardly, couldnt, wasnt, didnt,wouldnt, shouldnt, werent, dont, doesnt, havent, hasnt, wont, hadnt,never), the meaning of all succeeding words is inverted.

8

3. Stop word removal. Words without a deeper meaning, such as the, is,of, etc. are named stop words [42] and can be removed. We use a list of571 stop words proposed in [43].

4. Synonymmerging. Synonyms, though spelled differently, convey the samemeaning. In order to group synonyms by their meaning, we follow a methodthat is referred to as pseudoword generation [42]. Approximately 150 fre-quent synonyms or phrases from the finance domain are aggregated accord-ing to their meanings.

5. Stemming. Stemming refers to the process of reducing inflected words totheir stem [42]. Here, we use the so-called Porter stemming algorithm [44].

Having completed the preprocessing, we can continue to analyze news sen-timent. As shown in a recent study [45] on the robustness of sentiment anal-ysis, the correlation between news sentiment and stock market returns variesacross different sentiment metrics. A sentiment approach that results in a reli-able correlation is the Net-Optimism metric [17]. Net-Optimism works well incoordination with Henrys Finance-Specific Dictionary [9]. Consequently, werely upon this approach in the following evaluation. The metric is calculatedas the difference between the number of positive Wpos(A) and negative Wneg(A)words divided by the total number of words Wtot(A) in an announcement A.Thus, Net-Optimism sentiment S(A) is defined by

S(A) =Wpos(A)Wneg(A)

Wtot(A) [1,+1] . (1)

The relationship of news sentiment S(A) and corresponding stock market re-turns is depicted in Fig. 3. This diagram features a trend line from a so-calledgeneralized additive model (GAM), in which the linear predictor depends lin-early on unknown smooth functions of some predictor variables [46]. Fromthis GAM trend line, we can identify a visible relationship between the senti-ment variable and stock market returns.

9

Figure 3: GAM (generalized additive model) trend line showing relationship between newssentiment and stock market returns.

4. Trading Strategies

This section introduces all the trading strategies that serve as a foundationfor our analysis. Consistent with the existing literature, we start by presentingour benchmark, namely, a momentum trading approach. This strategy derivespurchase decisions solely from the historic returns of assets by maximizing theso-called rate-of-change. In addition, we propose news-based trading strate-gies in which investment decisions are triggered by news sentiment signals.Then, we combine both methods and develop a strategy that utilizes both his-toric prices and news sentiment.

In the subsequent algorithms, we use the following notation: let pi,t denotethe closing price of a stock i at time t. Furthermore, the variable S(A) givesthe news sentiment of an announcement A corresponding to stock i as definedabove.

When trading, we exclude all so-called penny stocks (i. e. stocks below 5) from our evaluation [47, 48]. The reason behind this is that these pennystocks tend to react more unsystematically to trends and news announcementsand, consequently, may introduce a larger noise component to our data.

10

4.1. Benchmark: Momentum TradingPast stock returns can be a predictor of future firm performance. This is

what we define as momentum, in which historic stock prices continue moving intheir previous direction. The (partly) predictable connection between past andfuture return has been proven in the finance literature, such as in [49]. Never-theless, finance academics have trouble with the finding that a simple strategyof buying winners and selling losers can apparently be profitable, since this con-tradicts the theory of efficient markets [2], where markets quickly absorb newinformation and adjust asset prices accordingly. Momentum is, consequently,also named a premier anomaly in stock returns [50]. By extrapolating his-toric stock trends, we motivate the following momentum trading, which picksup the subtle patterns in returns. Developing a successful momentum tradingstrategy is primarily a merit from the manual efforts of finance academics andpractitioners to hand-engineer features from historical prices [51].

Let us define both the terms momentum and rate-of-change respectively.The so-called momentum Momi,t is the absolute difference in stock i definedby

Momi,t = pi,t pi,t (2)with a time span of days. In short, momentum denotes the difference be-tween todays closing price and the closing price N days ago, thus referring toprices continuing to trend. In comparison, the rate-of-change RoCi representsthe relative change as a fraction, i. e.

RoCi,t =pi,t pi,t

pi,t=

Momi,tpi,t

. (3)

Both the momentum and rate-of-change indicators indicate trend by remainingpositive during an uptrend or negative during a downtrend.

Altogether, this results in the momentum trading strategy [52], formallyintroduced by the following pseudocode. In short, the key idea is to alwayschoose the stock that has the highest rate-of-change. Step 1 initializes thevariable s which stores the stock that our Decision Support System currentlyholds. The subequent for-loop iterates through all time steps of our simulationhorizon T . In each iteration, Step 3 updates the rate-of-change scales for allstocks, excluding the current business day. If the previously held stock wasempty, then Step 5 invests in the stock with the highest absolute value of allhistoric rate-of-change values. However, if the rate-of-change of the currentlyheld stock drops below a threshold RoC, then we trigger transactions to sell

11

the previous stock (Step 7) and buy (or short-sell) new stock with the highestrate-of-change in Step 8.

As free parameters, we can vary the time span calculating the rate-of-change and the threshold RoC. For the former, we find good results with set to 200 business days. This value serves as a good trade-off in between therange of 20 days to 12 months proposed in the literature [49]. We choose thelatter variable RoC by testing different values, and decide for RoC = 50%.

1: Initialize stock s.2: for t in T do

3: Compute RoCi,t1 pi,t1 pi,t1pi,t1 for all stocks i.4: if s = then5: Buy or short-sell stock s argmax

i|RoCi|.

6: else ifRoCs,t1< RoC then

7: Remove investment in stock s.8: Buy or short-sell stock s argmax

i|RoCi|.

9: end if10: end for

4.2. News TradingWhile the previous strategy only utilizes historic stock prices, we now in-

stead focus on news sentiment in order to enable news-based purchase deci-sions. In order to react to news sentiment signals, our Decision Support Systemneeds to continuously scan the news stream and compute the sentiment once anew financial disclosure is released. When the news sentiment associated withthis press release is either extremely positive or negative, this implies a stronglikelihood of a subsequent stock market reaction in the same direction. Webenefit from the stock market reaction if an automated transaction is triggeredbeforehand.

To achieve this goal, we specify the so-called simple news trading strat-egy, given in the pseudocode below. Steps 2 and 3 trigger buy and short-selldecisions, whenever the absolute value of the news sentiment metric of an in-coming announcement exceeds a certain positive or negative threshold. Thisdecision is given by the if-statement in Step 1, i. e. the condition that S(A) issmaller than a negative threshold S or larger than a positive

+S must be ful-

filled. We choose suitable threshold values for both S and +S as part of our

12

evaluation in Section 5.

Input: Released announcement A that corresponds to stock i.1: if S(A)> +S or S(A)<

S then

2: Remove investment in previous stock s.3: Buy or short-sell stock s i.4: end if

4.3. Combined Strategy with News and Momentum TradingThe subsequent trading strategy combines the above approaches by utiliz-

ing both news sentiment and historic prices in the form of momentum. Wedevelop this trading strategy around the idea that we want to invest in assetswith both (1) a news disclosure with a high polarity and (2) a previous mo-mentum in the same direction. The combined pseudocode is given below. Onlyif both the news release and historic prices give an indication of a developmentin the same direction, Steps 3 and 4 trigger a corresponding trading decision.Thus, this strategy expects the same direction in terms of the return-of-changeand sentiment metric as tested in Step 2.

Input: Released announcement A that corresponds to stock i at day t.

1: Compute RoCi,t1 pi,t1 pi,t1pi,t1 for stock i.2: if |S(A)|> S and sign S(A) = sign RoCi,t1 then3: Remove investment in stock s.4: Buy or short-sell stock s i.5: end if

5. Evaluation

The above sections have presented a number of trading strategies that differin the way in which operations are derived; this section evaluates these trad-ing strategies in terms of their achieved performance. We first focus on ourbenchmark strategies, i. e. momentum trading and German composite stockmarket index. We then determine suitable values for all free parameters insidethe news trading algorithms and analyze their performance.

5.1. Benchmarks: Stock Market Index and Momentum TradingAs our first benchmark, we choose the so-called CDAX, a German stock

market index calculated by Deutsche Brse. It is a composite index of all stocks

13

traded on the Frankfurt Stock Exchange that are listed in the General Standardor Prime Standard market segments. Figure 2 shows the overall developmentof the stock index from the year 2004 until mid-2011. During that period, theindex increased by 50.99 %, which correspond to 5.65 % at an annualized rate.The number of days with positive returns outweigh the negative days by 1092to 864.

Simple momentum trading acts as our second benchmark. This strategyworks with no data input other than historic stock prices. When historic pricescontinue their trend, we can invest in the specific stock to profit from thisdevelopment. While news trading yields high returns at first, the profits laterplummet, resulting in a negative cumulative return of65.33 %. Nevertheless,the average daily returns remain positive at 0.0464 % and are even higher thanthat of the CDAX index (0.0298 %). The number of days with positive returnsoutweighs the negative by 1154 to 767.

The results of both benchmarks, namely, the CDAX stock market index andmomentum trading, are presented in Fig. 4 where we see different performancepatterns. When looking at the histogram of returns in Fig. 5, we see that volatil-ity is smaller for the CDAX. The volatility of daily returns accounts for 0.013 forthe CDAX index, while it is higher at 0.045 in the case of momentum trading.

-30

0

30

60

Jan 2004 Apr 2004 Jul 2004 Oct 2004 Jan 2005 Apr 2005 Jul 2005

Cumu

lative

Retur

n (in

%)

CDAX Stock IndexMomentum Trading

Figure 4: Cumulative returns of both the CDAX index and the momentum trading strategycompared across the first 400 business days.

14

0100

200

300

-4 0 4Daily Return (in %)

Frequ

ency

CDAX Stock Index

050

100150200

-20 -10 0 10 20Daily Return (in %)

Frequ

ency

Momentum Trading

Figure 5: Histogram of returns for the CDAS index (left) and the momentum trading strat-egy (right), in which the vertical bars denote the corresponding mean value.

5.2. News TradingThis section evaluates different variants of news trading, starting with a

simple trading strategy. This strategy triggers transactions whenever a verypositive or negative ad hoc announcement is disclosed. Here, we take intoaccount only a single announcement per day and business days, giving a totalcorpus of 1894 disclosures.

What remains unanswered thus far is av value for the threshold S abovewhich our news-based trading strategies perform a purchase decision. In orderto find the optimal parameter, Fig. 6 compares the thresholds +S and

S against

the average returns. For reasons of simplicity, we measure these thresholds interms of quantiles of the news sentiment distribution. We see that a thresholdvalue of around 10 % appears in a cluster of high daily returns and yields goodresults. Thus, we decided to set +S to the 90 % quantile (and

S to the 10 %

quantile) of the sentiment values S(A) in order to make this variable exoge-nously given. However, it is important to stress that there large variations inperformance depending on the threshold.

Evaluating the above strategies with historic data reveals the following find-ings:

With the threshold set to the 10 % quantile, we gain an overall cumulativereturn of 178.53 % at a volatility of 0.078. The average daily return accountsfor 0.421 %.

In addition, we include a combination of news and momentum trading. Thisstrategy leads to a lower performance with a cumulative return of 361.60 %and average daily returns of 0.1180 %. However, this strategy simultane-

15

ously reveals a reduced risk component in the form of less volatility, whichaccounts for 0.028.

Both strategies, namely, simple news trading and the combined version, arefurther evaluated in the following diagrams. Fig. 7 depicts how the cumula-tive returns develop during the first 1500 business days, showing clearly thesuperiority of simple news trading. Analogously, Fig. 8 compares the distribu-tion of daily returns. We note that evidently higher volatility of simple newstrading, which is ultimately linked to higher risks.

0.3

0.4

0.5

0.6

10 20 30 40 50Threshold _S as Quantile of News Sentiment (in %)

Avera

ge Da

ily Re

turn (

in %)

Cumulative Return (in %)05001500

Figure 6: Comparison of thresholds +S and S against average daily returns. The threshold

is measured as quantiles from both ends of the average news sentiment in the corpus. Inaddition, the point size indicates the total cumulative return.

16

02000

4000

6000

2004 2005 2006 2007 2008 2009 2010

Cumu

lative

Retur

n (in

%) Combined News & Momentum TradingNews Trading

Figure 7: Cumulative returns of both news trading and the combination of news and momen-tum trading compared across the first 1500 business days.

0

100

200

300

-15 -10 -5 0 5 10 15Daily Return (in %)

Frequ

ency

News Trading

0

100

200

300

-15 -10 -5 0 5 10 15Daily Return (in %)

Frequ

ency

Combined News&Momentum Trading

Figure 8: Histogram of returns comparing simple news trading (left) and the combined news& momentum trading strategy (right), in which the vertical bars denote the correspondingmean value.

5.3. ComparisonThe simulation horizon spans January 2014 until the end of June 2011,

giving a total of 1956 business days. All results of our trading simulation areprovided in Table 1. Here, we evaluate how well the investment decisions ofeach strategy play together with market feedback. We focus mainly on aver-age daily return, since cumulative returns can be misleading. The reason is as

17

follows: one wrong trade can make performance plummet, while a high aver-age daily return indicates a continuously high benefit. In addition, we wantto direct attention to the volatility column. These values serve as an indicatorof the level of risk associated with each strategy. Even though simple newstrading achieves higher returns, it is linked with higher volatility and higherrisks. Thus, it may be beneficial for practitioners to follow a combined strat-egy of news and momentum trading that results in smaller returns, while alsodecreasing the associated risk.

We now compare our benchmarks to news trading. The benchmarks featuremean returns of 0.0298 % for the CDAX and 0.0464 % for momentum trading.In comparison, news trading reaches 0.4206 %. This is the 9.06-fold value, butlinked with a considerably higher volatility.

While we put an emphasis on raw returns, we also provide a performancemeasure that incorporates transaction costs. Thus, the last column of Table 1reveals how an initial invest of 1000 would evolve over time. Consistentwith [29, 8]3, we simulate the portfolio with a transaction fee for each buy/selloperation of 0.1 %, equivalent to 10 bps.4

3Using a proportional transaction fee is common in financial research. For example, otherpapers [30] mostly vary transaction costs mostly in the range of 0.1 % to 0.3 % or assume afixed transaction fee [33] of U. S. $ 10 for buying and selling stocks respectively.

4A frequent unit in finance is basis point (bps). Here, one unit is equal to 1/100th of 1 %,i. e. 1 % = 100 bps.

18

Table 1: Comparison of benchmarks and trading strategies across several key performancecharacteristics.

TradingStrategy

Cumulative

Return

Returns:

Median

Returns:

Mean

Annualized

Return

#Positive

Returns

#Negative

Returns

Volatility

Trades

Portfolio

Outcom

e(

)

Benchmarks

CD

AX

Inde

x50

.99

%0.

0703

%0.

0298

%5.

6480

%10

9286

40.

013,

19

1506

.92

Mom

entu

mTr

adin

g

65.3

3%

0.00

00%

0.04

64%

13

.170

2%

1129

827

0.04

4,96

102.

95d

322.

94(

=50

%,

=20

0d)

New

sTrading

Sim

ple

New

sTr

adin

g17

,853

.31

%0.

0000

%0.

4206

%99

.780

3%

1136

820

0.07

7,80

10.5

7d

967,

448.

60(

+ S=

90%

, S=

10%

)C

ombi

ned:

New

s&

Mom

entu

m36

1.60

%0.

0000

%0.

1180

%22

.622

2%

1113

843

0.02

8,39

19.9

6d

9513

.92

(+ S=

90%

, S=

10%

,=

200

d)

19

CDAX IndexMomentum Trading

Simple News Trading

News Trading [Chemicals]

News Trading [Automobile]Combined News & Momentum Trading

News Trading with Min. Holding Time

0.1

0.2

0.3

0.4

0.000 0.025 0.050 0.075 0.100 0.125Volatility of Daily Returns

Avera

ge Da

ily Re

turn (

in %)

Figure 9: risk-return-comparison.

5.4. Statistical TestsThis section provides statistical evidence that trading strategies can result in

daily returns above zero. We perfom both a one-sided Wilcoxon test and a one-sided t-test. In both cases, the null hypothesis tests whether a given tradingstrategy results in a zero mean value of daily returns. All test outcomes areprovided in the form of P-values in Table 2. Apparently, a few P-values arestatistically significant at common significance levels. Consequently, in thesecases, we can reject the null hypothesis at the corresponding significance leveland conclude that the trading strategies have an above zero mean of dailyreturns. This is particulary true for simple news trading, providing that newstrading yields statistically significant returns.

20

Table 2: Statistical tests to validate the hypothesis that daily returns are above zero.

Trading Strategy Daily ReturnsMean

Wilcoxon TestOne-Sided(P-Value)

t-TestOne-Sided(P-Value)

BenchmarksCDAX Index 0.0298 % 0.0031 0.1593Momentum Trading 0.0464 % 0.7803 0.3241

News TradingSimple News Trading 0.4206 % 0.0038 0.0085Combination News & Momentum 0.1180 % 0.0474 0.0031

Significance levels: 0.001, 0.01, 0.05

5.5. Discussion and Managerial ImplicationsThis paper aims to develop a core functionality of an automated system for

triggering news-based trading decisions. Although it is part of an algorithmic-trading system, the system is often regarded as a Decision Support System [10].In the above evaluation, such a system demonstrated that news trading is moreprofitable overall than the benchmark scenarios. In practice [10], such a De-cision Support System may consist of 10 to 100 data/text mining algorithmsand, in addition, model parameters must frequently be calibrated to the latestdatasets. These text mining algorithms have contributed to the reduction ofmarket noise from news by extracting only the relevant events from the overallnews pile [53].

One of the challenges remaining is that of the information quality of news.Although news plays a significant role in driving stock prices, its impact is low,leaving a large portion of noise [8]. This inevitably results in unpredictableand possibly reduced market reactions when signals are unclear [54]. Differentdata sources might provide further insights into the costs of additional noise.For example, innovative news sources [10] offer a massive number of newcorpus entries, which are published quickly without quality control, such associal media like Twitter. Related approaches also incorporate Google queryvolume for search terms related to finance [34] and Internet stock messageboards [6]. Each of these sources has its own advantages, but practitioners arelikely to implement a combination. Regardless, we need a remedy that worksreliably even when facing noisy data. To overcome this challenge, we proposednews-trading strategies in the form of rule-based algorithms.

The above research opens up several other research streams. The proposedDecision Support System with its trading strategies can also be highly relevant

21

for high-frequency trading [55] and provides valuable insights for this adjacentdiscipline, relying purely on automated trading algorithms. News trading alsointersects with portfolio management problems [56, for example], in whichinvestors make decisions about which portfolio to hold at an ex-ante basiswithout knowing future returns. Though these returns are unknown, mar-ket participants make decisions and trigger transactions based on their futureexpectations of the market.

Managers on the road to implementing such news trading systems canchoose between a vast number of options. However, we have seen that func-tional verifications are rare. It is, therefore, no surprise that news trading re-mains highly specialized and continues to be based on disparate decision-makingtools and analytical models [10].

6. Conclusion and Outlook

Although it is a well-known fact that financial markets are very sensitiveto the release of financial disclosures, the way in which this information is re-ceived is far from being studied sufficiently. Not until recently have researchersstarted to look at the content of news stories using very simple techniques todetermine news sentiment. Typically, these research papers concentrate onfinding a link between the qualitative content and the subsequent stock mar-ket reaction. To harness this relationship in practice, news trading combinesreal-time market data and sentiment analysis in order to trigger investmentdecisions. Interestingly, what previous approaches all have in common is thatthey rarely study and compare trading strategies.

As a remedy, this paper evaluates algorithmic trading strategies within aDecision Support System for news trading. We propose and compare differentrule-based strategies for news-based trading. As a result, our Decision Sup-port System outperforms all benchmark scenarios by relying upon news-basedinvestment decisions. Further performance improvements can be achieved byincluding external variables that, for example, describe the economic environ-ment or lagged prices respectively. We also find considerable differences intrading performance across firm sectors, with the automobile and chemicalsectors yielding an above average stock market return. The reasons for thisseem to be that these domains are more reactive to novel announcements andthe expected return direction more reliable. Altogether, we contribute to theunderstanding of information processing in electronic markets and show howto enable decision support in financial markets.

22

This paper opens avenues for further research into two directions. First, amulti-asset strategy could be beneficial to spread risks. To model these, intrigu-ing approaches include Value-at-Risk (VaR) measures, as well as techniquesfrom portfolio optimization. Second, it is worthwhile to improve the forecastof asset returns by including a broader set of exogenous predictors. As such,possible external variables range from stock market indices, fundamentals de-scribing the economy and additional lagged variables. Further enhancementswould also result from embedding innovative news sources, such as social me-dia. Altogether, the accuracy of predicting stock return directions and triggertrading signals would be greatly improved.

Acknowledgment

The valuable contributions of Dirk Neumann, Laura Cuthbertson and Si-mon Schmidt are gratefully acknowledged.

[1] N. Granados, A. Gupta, R. J. Kauffman, Research CommentaryInformation Transparency in Business-to-Consumer Markets: Concepts,Framework, and Research Agenda, Information Systems Research 21(2010) 207226.

[2] F. E. Fama, The Behavior of Stock-Market Prices, The Journal of Business38 (1965) 34105.

[3] T. Cenesizoglu, The Reaction of Stock Returns to News About Fundamen-tals, Management Science (2014).

[4] D. M. Cutler, J. M. Poterba, L. H. Summers, What Moves Stock Prices?,Journal of Portfolio Management 15 (1989) 412.

[5] R. J. Shiller, Irrational Exuberance, 2nd ed., Princeton University Press,2005.

[6] W. Antweiler, M. Z. Frank, Is All That Talk Just Noise? The InformationContent of Internet Stock Message Boards, Journal of Finance 59 (2004)12591294.

[7] P. C. Tetlock, Giving Content to Investor Sentiment: The Role of Mediain the Stock Market, Journal of Finance 62 (2007) 11391168.

23

[8] P. C. Tetlock, M. Saar-Tsechansky, S. Macskassy, More than Words: Quan-tifying Language to Measure Firms Fundamentals, Journal of Finance63 (2008) 14371467.

[9] E. Henry, Are Investors Influenced By How Earnings Press Releases AreWritten?, Journal of Business Communication 45 (2008) 363407.

[10] S. Gagnon, Rules-Based Integration of News-Trading Algorithms, Journalof Trading 8 (2013) 1527.

[11] R. P. Schumaker, H. Chen, Textual Analysis of Stock Market Prediction us-ing Breaking Financial News, ACM Transactions on Information Systems27 (2009) 119.

[12] R. P. Schumaker, H. Chen, A Quantitative Stock Prediction System basedon Financial News, Information Processing & Management 45 (2009)571583.

[13] R. P. Schumaker, Y. Zhang, C.-N. Huang, H. Chen, Evaluating Sentimentin Financial News Articles, Decision Support Systems 53 (2012) 458464.

[14] B. Pang, L. Lee, Opinion Mining and Sentiment Analysis, FNT in In-formation Retrieval (Foundations and Trends in Information Retrieval) 2(2008) 1135.

[15] M.-A. Mittermayer, G. F. Knolmayer, Text Mining Systems for Market Re-sponse to News: A Survey, 2006.

[16] M. Minev, C. Schommer, T. Grammatikos, News and Stock Mar-kets: A Survey on Abnormal Returns and Prediction Models,2012. URL: http://publications.uni.lu/record/9947/files/TR.Survey.News.Analytics.pdf.

[17] E. A. Demers, C. Vega, Soft Information in Earnings Announcements:News or Noise? INSEAD Working Paper No. 2010/33/AC, SSRN Elec-tronic Journal (2010).

[18] N. Jegadeesh, D. Wu, Word Power: A New Approach for Content Analy-sis, Journal of Financial Economics 110 (2013) 712729.

24

[19] T. Loughran, B. McDonald, When Is a Liability Not a Liability? TextualAnalysis, Dictionaries, and 10-Ks, Journal of Finance 66 (2011) 3565.

[20] F. Li, The Information Content of Forward-Looking Statements in Corpo-rate Filings: A Narve Bayesian Machine Learning Approach, Journal ofAccounting Research 48 (2010) 10491102.

[21] M.-A. Mittermayer, G. F. Knolmayer, NewsCATS: A News Categorizationand Trading System, in: Sixth International Conference on Data Mining(ICDM06), 2006, pp. 10021007. doi:10.1109/ICDM.2006.115.

[22] A. Sharma, S. Dey, A Comparative Study of Feature Selection and Ma-chine Learning Techniques for Sentiment Analysis, in: Y. Cho, A. S. I.G. o. A. Computing (Eds.), Proceedings of the 2012 Research in AppliedComputation Symposium (RACS 2012), ACM, New York, NY, 2012, pp.17. doi:10.1145/2401603.2401605.

[23] N. Barberis, R. Thaler, A Survey of Behavioral Finance, in: Finan-cial Markets and Asset Pricing, volume 1 of Handbook of the Economicsof Finance, Elsevier, Amsterdam, Netherlands, 2003, pp. 10531128.doi:10.1016/S1574-0102(03)01027-6.

[24] A. Shleifer, L. H. Summers, The Noise Trader Approach to Finance, TheJournal of Economic Perspectives 4 (1990) 1933.

[25] A. Shleifer, R. W. Vishny, The Limits of Arbitrage, Journal of Finance 52(1997) 3555.

[26] S. J. Alfano, S. Feuerriegel, D. Neumann, Is News Sentiment More thanJust Noise?, SSRN Electronic Journal (2014).

[27] B. Mendel, A. Shleifer, Chasing Noise, Journal of Financial Economics104 (2012) 303320.

[28] F. Li, Do Stock Market Investors Understand the Risk Sentiment of Cor-porate Annual Reports?, SSRN Electronic Journal (2006).

[29] M. Hagenau, M. Liebmann, M. Hedwig, D. Neumann, Automated NewsReading: Stock Price Prediction Based on Financial News Using Context-Specific Features, in: 45th Hawaii International Conference on SystemSciences (HICSS), 2012, pp. 10401049. doi:10.1109/HICSS.2012.129.

25

[30] F. Graf, Mechanically Extracted Company Signals and theirImpact on Stock and Credit Markets, 2011. URL: http://www.uni-konstanz.de/FuF/wiwi/workingpaperseries/

WP\textunderscore18-11-Graf.pdf.

[31] F. M. F. Wong, Z. Liu, M. Chiang, Stock Market Prediction from WSJ:Text Mining via Sparse Matrix Factorization, in: IEEE International Con-ference on Data Mining (ICDM), IEEE Computer Soc., Shenzhen, China,2014, pp. 430439.

[32] A. Klein, O. Altuntas, T. Hausser, W. Kessler, Extracting Investor Senti-ment from Weblog Texts: A Knowledge-based Approach, in: IEEE 13thConference on Commerce and Enterprise Computing (CEC), 2011, pp.19. doi:10.1109/CEC.2011.10.

[33] M.-A. Mittermayer, Forecasting Intraday stock price trends with text min-ing techniques, in: R. H. Sprague (Ed.), Proceedings of the 37th AnnualHawaii International Conference on System Sciences, IEEE ComputerSoc., Los Alamitos, Calif., 2004. doi:10.1109/HICSS.2004.1265201.

[34] T. Preis, H. S. Moat, H. E. Stanley, Quantifying Trading Behavior in Fi-nancial Markets using Google Trends, Scientific Reports 3 (2013) 1684.

[35] S. Feuerriegel, A. Ratku, D. Neumann, Finding Evidence of IrrationalExuberance in the Oil Market, SSRN Electronic Journal (2014).

[36] S. Feuerriegel, A. Ratku, D. Neumann, Which News Disclosures Mat-ter? News Reception Compared Across Topics Extracted from the LatentDirichlet Allocation, SSRN Electronic Journal (2015).

[37] J. Muntermann, A. Guettler, Intraday Stock Price Effects of Ad Hoc Dis-closures: The German Case, Journal of International Financial Markets,Institutions and Money 17 (2007) 124.

[38] S. Feuerriegel, S. F. Heitzmann, D. Neumann, Do Investors Read TooMuch into News? How News Sentiment Causes Price Formation, in: 48thHawaii International Conference on System Sciences (HICSS), Forthcom-ing, 2015.

26

[39] G. Grefenstette, P. Tapanainen, What Is a Word, What Is a Sentence?Problems of Tokenization, 1994. URL: http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.28.5162.

[40] M. Dadvar, C. Hauff, F. de Jong, Scope of Negation Detection in Sen-timent Analysis, in: Proceedings of the Dutch-Belgian Information Re-trieval Workshop (DIR 2011), Amsterdam and Netherlands, 2011, pp.1620.

[41] N. Prullochs, S. Feuerriegel, D. Neumann, Enhancing Sentiment Analy-sis of Financial News by Detecting Negation Scopes, in: 48th Hawaii In-ternational Conference on System Sciences (HICSS), Forthcoming, 2015,pp. 959968.

[42] C. D. Manning, H. Schijtze, Foundations of Statistical Natural LanguageProcessing, MIT Press, Cambridge and Mass, 1999.

[43] D. D. Lewis, Y. Yang, T. G. Rose, F. Li, RCV1: A New Benchmark Collectionfor Text Categorization Research, Journal of Machine Learning Research5 (2004) 361397.

[44] M. F. Porter, An Algorithm for Suffix Stripping, Program 14 (1980) 130137.

[45] S. Feuerriegel, D. Neumann, News or Noise? How News Drives Commod-ity Prices, in: Proceedings of the International Conference on InformationSystems (ICIS 2013), Association for Information Systems, 2013.

[46] T. J. Hastie, R. J. Tibshirani, J. H. Friedman, The Elements of Statisti-cal Learning: Data Mining, Inference, and Prediction, Springer Series inStatistics, 2nd ed., Springer, New York, NY, 2013.

[47] R. Beatty, P. Kadiyala, Impact of the Penny Stock Reform Act of 1990 onthe Initial Public Offering Market, The Journal of Law and Economics 46(2003) 517541.

[48] E. F. Fama, K. R. French, Multifactor Explanations of Asset Pricing Anoma-lies, Journal of Finance 51 (1996) 5584.

[49] N. Jegadeesh, S. Titman, Returns to Buying Winners and Selling Losers:Implications for Stock Market Efficiency, Journal of Finance 48 (1993)6591.

27

[50] E. F. Fama, K. R. French, Dissecting Anomalies, Journal of Finance 63(2008) 16531678.

[51] L. Takeuchi, Y.-Y. Lee, Applying Deep Learning to Enhance MomentumTrading Strategies in Stocks, 2013.

[52] K.-j. Kim, Financial Time Series Forecasting using Support Vector Ma-chines, Neurocomputing 55 (2003) 307319.

[53] A. Gro-Klumann, N. Hautsch, When Machines Read the News: Us-ing Automated Text Analytics to Quantify High Frequency News-ImpliedMarket Reactions, Journal of Empirical Finance 18 (2011) 321340.

[54] P. C. Tetlock, Does Public Financial News Resolve Asymmetric Informa-tion?, Review of Financial Studies 23 (2010) 35203557.

[55] I. Aldridge, High-Frequency Trading: A Practical Guide to AlgorithmicStrategies and Trading Systems, Wiley Trading Series, 2nd ed., Wiley,Hoboken, NJ, 2013.

[56] L. Mitra, G. Mitra, D. Dibartolomeo, Equity Portfolio Risk Estimationusing Market Information and Sentiment, Quantitative Finance 9 (2009)887895.

28