Predicting Market Movements: From Breaking News to Emerging Social Media Dr. Hsinchun Chen Director, Artificial Intelligence Lab University of Arizona hchen@eller.arizona.edu http://ai.arizona.edu Acknowledgements: NSF CRI; NSF EXP-LA; DOD DTRA, CTFP, NPS; (ARFL WMD, CIA, FBI) PREDICITNG MARKET MOVEMENTS Predicting Markets Markets: international markets, emerging markets, import/export markets, financial market, stock market, commodity market, retail market Economics (macro), international relations (trade, geopolitics), finance (international/banking/stock), accounting (market return), marketing (sales/retailing) US (NSF SBE, social behavioral economics; governments, think tanks), Europe/Asia Business school research in not science (cannot be funded by NSF in US)! Economics, finance, accounting, political science, social science, marketing, computer science (small, no funding in US!), MIS (business intelligence) Geopolitical/econ/finance/accounting models/theories, market metrics/parameters, analytical techniques, results interpretations, predicating markets EMH (efficiency market hypothesis), RWT (random walk theory), CAPM (capital asset pricing model), quant/algorithm trading Research Opportunities Sophisticated econ/finance/accounting/marketing models/theories, established analytical techniques and metrics (numeric), abundant structured databases (financial metrics, economic indicators, stock quotes) New, diverse unstructured (text) web-enabled business data sources, e.g., 10K/10Q SEC reports, mass media news, local news, Internet news, financial blogs, investor forums, tweets… Topic extraction, named entity recognition, sentiment/affect analysis, multilingual language models, social network analysis, statistical machine learning, temporal data/text mining, timeseries analysis… Nerds on Wall Street “Future technological stars…(1) Advanced electronic market tools; (2) Understanding both quantitative and qualitative information…” “The Text Frontier, Collective Intelligence, Social Media, and Market Monitors” “Stocks are stories, bonds are mathematics.” David Leinweber, 2009 AZ BIZ INTEL: BUSINESS MASS MEDIA, SOCIAL MEDIA, TEXT ANALYTICS, SENTIMENT ANALYSIS, SPIKE DETECTION, FINANCE/ACCOUNTING/MARKETING MODELING, PREDICTING MARKET MOVEMENTS Business Intelligence & Analytics • • • • $3B BI revenue in 2009 (Gartner, 2006) The Data Deluge (The Economists, March 2010); internet traffic 667 Exabytes by 2013, Cisco; Total amount of information in 2010, 1.2 Zettabyte (KB-MB-GB-TB-PB-EB-ZBYB) $9.4B BI software M&A spending in 2010 and $14.1B by 2014 (Forrester) IBM spent $14B in BI in five years; $9B BI revenue in 2010 (USA Today, November 2010); 24 acquisitions, 10,000 BI software developers, 8,000 BI consultants, 200 BI mathematicians Acquired i2/COPLINK in 2011 Business Intelligence & Analytics • BI: “skills, technologies, applications, and practices used to help an enterprise better understand its business and market.” • Technologies: data warehousing; Extraction, Transformation, and Load(ETL); Business Performance Management (BPM); visual dashboards; and advanced knowledge discovery using data and text mining BI 2.0: web intelligence, web analytics, web 2.0, social media analytics, opinion mining; cloud computing and web services; real-time monitoring and mining; enterprise performances (marketing/accounting/finance/healthcare) • AZ BIZ INTEL • • • • • • • Mass media, social media contents Text & social media analytics techniques Finance/accounting/marketing models (Tetlock/Columbia, Antweiler/UBC, Das/Santa Clara) NYU (Dhar), Arizona (Dhaliwal, Kelly, Jiang, Lusch, Yong), National Taiwan U (Li, Hong, Lu) Bag of words, named entities, proper nouns, topics (1, 2-, 3- grams) Sentiment/valence, lexicons, machine learning, stakeholder analysis, EFLS analysis Time series models, spike detection, decaying function, trading windows, targeted sentiment Econometrics/regression models (R-sqr, p-value), 10-fold validation (F, accuracy), simulated trading (cost, frequency, exit) AZ ONLINE WOM AZ WOM: events, volume, sentiment Data Collection Yahoo! Movie Parsing Data Processing OpinionFinder SentiWordNet Sales Data Professional Evaluation Firms Strategy Online WOM evolution Correlation between different WOM measures Measures and Metrics Online WOM measures Messages Statistical Analysis Number of messages Number of sentences Valence Subjectivity Number of valence words New-product performance metrics Opening-week box office sales Total box office sales Opening strength Longevity Professional evaluation Correlation of WOM measure across newproduct lifecycle Correlation between online WOM and product performance Correlation between online WOM measures and new-product performance across the whole new-product lifecycle 11 Results Evolution of online WOM through new-product lifecycle WOM communication starts early in preproduction, becomes highly active before movie release, then diminishes gradually Valence has a clear decreasing trend over time, indicating that WOM becomes more negative after movie release Subjectivity, number of sentences and number of valence words stay stable over time 12 IT’S THE BUZZ! 13 AZ STOCK TRACKER I & II Literature Review: Stock Performance Prediction Theoretical perspectives on stock behavior Efficient market hypothesis (Fama 1964) Random walk theory (Malkiel 1973) Price of a stock reflects all available information Market reacts instantaneously; impossible to outperform Price of a stock varies randomly over time Future prediction, outperforming the market is impossible Pessimistic assessments of the predictability of stock behavior refuted through empirical studies Lo and MacKinlay 1988; Jaffe et al 1989; Pesaran and Timmermann 1995 15 Literature Review: Stock Performance Prediction Predominant approaches to stock prediction Fundamentalists utilize fundamental and financial measures of economy, industry, and firm Economy and sector indicators, financial ratios of the firm Technicians utilize historical time-series information of the stock and market behavior Fama-French three factors model (Fama and French 1993) Market return, market capitalization, book to market ratio Currency exchange rates, interest rates, dividends Historical price, volatility, trading volume Various machine learning models applied Regression, ANN, ARIMA, support vector machines 16 Literature Review: Stock Performance Prediction In addition to financial and stock variables, researchers have incorporated firm-related news article measures Developed trend-based language models for news articles Categorized press releases (good, bad, neutral) Mittermayer 2004 Examined various textual representations of news articles Lavrenko et al. 2000 Schumaker and Chen, 2009a; 2009b But few have incorporated firm-related web forums Thomas and Sycara (2000) utilize text classifications of discussions on Raging Bull to inform stock trading strategies 17 Literature Review: Firm-Related Web Forums and Stock Studies relating web forums and stock behavior Examined firm-related web forums on major web portals Early studies focused on activity, without content analysis Supported market efficiency; only concurrent relationships identified Subsequently challenged; forum activity predicted stock behavior Wysocki 1998; Tumarkin and Whitelaw 2001 Antweiler and Frank 2002; 2004; Das and Chen 2007 Analysis advanced to measure opinions in discussions ‘Bullishness’ classifiers to distinguish investment positions Antweiler and Frank 2004; Das and Chen 2007 Classified buy, hold, or sell positions with 60 – 70% accuracy Identified predictive relationships between forum discussion sentiment and subsequent stock returns, volatility, trading volume Shortcomings Retrospective analyses, shareholder perspective of major forums 18 AZ FinText: numbers + text • Techniques: bag of words, named entities, proper nouns, past stock prices + SVR • Testbed: S&P 500 5 weeks, Oct-Nov 2005, 2,809 news, 10M stock quotes, GICS industry classification • Evaluation: Return, vs. Quant funds; 20-minute prediction AZ FinText in the news Thursday, June 10, 2010 AI That Picks Stocks Better Than the Pros A computer science professor uses textual analysis of articles to beat the market. WSJ Technology News and Insights June 21, 2010, 1:45 PM ET Using Artificial Intelligence to Digest News, Trade Stocks AZ STOCK TRACKER I: mass, social media, topic, volume, sentiment Web Forums Mutual information phrase extractor Conversation analysis Traffic dynamics Topic correlation and evolution Sentiment identification Sentiment grader Database Author Discussion topics Spider/ Parser Topic Sentiment correlation and evolution Active topics and sentiments Sentiment aggregator Market prediction Message sentiments Message 21 t Online news Topic extraction Sentimen Data collection User-Generated Contents (UGC): Conversations of 30,000 Wal-Mart Constituents and 500,000 Responses Data sources Duration # of Threads # of Messages # of Users Wall Street Journal - WalMart-related News (WSJ) Aug 1999 - Mar 2007 N/A 4,081 657 Yahoo! Finance - WalMart Message Board (YAHOO) Jan 1999 - Jun 2008 139,062 441,954 25,500 Walmart-blows Forum - Employee Department Board (EMP) Dec 2003 - Oct 2008 7,440 102,240 2,930 Walmart-blows Forum - WalMart Sucks Board (WSB) Nov 2003 - Nov 2008 1,354 19,624 1,855 Wakeupwalmart Forum - General WalMart Discussion Board (GDB) Aug 2005 - Nov 2008 2,136 23,940 967 22 320 16000 280 14000 240 12000 200 10000 160 8000 120 6000 80 4000 40 2000 0 WSJ # of messages # of news Post Dynamics YAHOO EMP WSB GDB 0 99 00 01 02 03 04 05 Year 06 07 08 23 Average sentiment Sentiment Trend 0.01 0 WSJ YAHOO EMP WSB GDB -0.01 -0.02 -0.03 -0.04 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 Year 3 months' moving average sentiment 0.01 0 -0.01 -0.02 -0.03 YAHOO WSJ EMP WSB GDB -0.04 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 Year 24 Market Modeling Correlation Return Return 1 Volatility 0.0348 Volatility 1 Trading Volume Sentiment Trading Volume 1 0.0338 Disagreement -0.0507 -0.03578 Message Volume -0.3186 0.3131 Message Length 0.0473 -0.1840 Disagreement One Day Lag -0.0527 -0.0475 Message Volume One Day Lag -0.3433 0.3026 Message Length One Day Lag 0.0859 -0.1795 Subjectivity Sentiment One Day Lag Subjectivity One Day Lag -0.0425 Correlation coefficients with p<0.10 are shown (two-tailed test) Correlation Sentiment expressed in the forum contemporaneously correlates significantly with stock return Disagreement, volume, and length expressed in the forum also hold significant correlations with volatility and trading volume 25 Market Predictive Results (cont’d) Overall Forum Markett Sentimentt-1 Disagreementt-1 Message Volumet-1 Message Lengtht-1 Subjectivityt-1 Returnt 0.8723*** (31.33) 0.0025 (0.31) 0.0000 (0.04) -0.0007** (-2.29) 0.0002 (1.42) 0.0015 (1.46) Volatilityt -0.0010 (-0.25) 0.0074 (0.47) -0.0023*** (-4.94) -0.0122*** (-19.09) 0.0030*** (7.82) 0.0149*** (7.27) Trading Volumet 0.7627*** (15.06) -0.4275** (-2.06) 0.0140** (2.29) 0.1957*** (23.18) -0.0668*** (-13.24) -0.3014*** (-11.11) Note: *p<0.10;**p<0.05;***p<0.01 • • Predictive regression (t-1) The significant measures of forum discussions identified in contemporaneous regressions maintain their significance in the predictive regression models Additionally, sentiment expressed in the web forum holds a significant relationship with the trading volume on the following day • Positive sentiment reduces trading volume; negative sentiment induces trading activity 26 AZ STOCK TRACKER II: stakeholder analysis 27 Experimental Design: Description of Prediction Models Variables Description Dependent: RETURN t Stock return on day t (log difference of share price) Fundamental: FFSIZE FFBTM FFMARKET t-1 FFMARKET t-2 Technical: Fama-French firm size (prior year; market capitalization = share price * shares outstanding) Fama-French book-to-market ratio (prior year; book value / market value of shares) Fama-French market return on day t – 1 (log difference of S&P 500 index price) Fama-French market return on day t – 2 (log difference of S&P 500 index price) RETURN t-1 RETURN t-2 VOLATILITY t-1 VOLATILITY t-2 VOLUME t-1 VOLUME t-2 DAY d t Stock return on day t – 1 (log difference of share price) Stock return on day t – 2 (log difference of share price) Stock price volatility on day t – 1 (volatility modeled using a GARCH(1,1)) Stock price volatility on day t – 2 (volatility modeled using a GARCH(1,1)) Stock trading volume on day t – 1 (in log) Stock trading volume on day t – 2 (in log) Dummy variables for trading day of the week on day t t = days (t = 1, 2, …, n); day of the week (d = 1, …, 4) 28 Experimental Design: Description of Prediction Models Variables Description Forum: MESSAGES t-1 LENGTH t-1 SENTI t-1 VARSENTI t-1 SUBJ t-1 VARSUBJ t-1 Stakeholder: Number of messages posted in the forum on day t – 1 (in log (1 + messages)) Average length of messages posted in the forum on day t – 1 (in number of sentences) Average sentiment of messages posted in the forum on day t – 1 Variance in sentiment of messages posted in the forum on day t – 1 Average subjectivity of messages posted in the forum on day t – 1 Variance in subjectivity of messages posted in the forum on day t – 1 MESSAGES s t-1 LENGTH s t-1 SENTI s t-1 VARSENTI s t-1 SUBJ s t-1 VARSUBJ s t-1 Number of messages posted by stakeholder cluster s on day t – 1 (in log (1 + messages)) Average length of messages posted by stakeholder cluster s on day t – 1 (in number of sentences) Average sentiment of messages posted by stakeholder cluster s on day t – 1 Variance in sentiment of messages posted by stakeholder cluster s on day t – 1 Average subjectivity of messages posted by stakeholder cluster s on day t – 1 Variance in subjectivity of messages posted by stakeholder cluster s on day t – 1 t = days (t = 1, 2, …, n); stakeholder clusters (s = 1, 2, …, c) 29 Experimental Design: Description of Prediction Models Baseline Model – Baseline-FF Fundamental variables: Fama-French model RETURN t = β0 + β1 FFSIZE + β2 FFBTM + β3 FFMARKET t-1 + β4 FFMARKET t-2 + εt Baseline Model – Baseline-Tech Technical variables: Lagged stock returns, volatility, trading volume, day-of-week dummies RETURN t = β0 + β1 RETURN t-1 + β2 RETURN t-2 + β3 VOLATILITY t-1 + β4 VOLATILITY t-2 + β5 VOLUME t-1 + β6 VOLUME t-2 + (β7 DAY1t + … + β10 DAY4t)+ εt Baseline Model – Baseline-Comp Comprehensive: all fundamental and technical variables RETURN t = β0 + β1 FFSIZE + β2 FFBTM + β3 FFMARKET t-1 + β4 FFMARKET t-2 + β5 RETURN t-1 + β6 RETURN t-2 + β7 VOLATILITY t-1 + β8 VOLATILITY t-2 + β9 VOLUME t-1 + β10 VOLUME t-2 + (β11 DAY1t + … + β14 DAY4t) + εt Where t = days (t = 1, 2, …, n); day of the week (d = 1, …, 4) 30 Experimental Design: Description of Prediction Models Forum models Comprehensive baseline variables plus forum-level measures RETURN t = β0 + β1 FFSIZE + β2 FFBTM + β3 FFMARKET t-1 + β4 FFMARKET t-2 + β5 RETURN t-1 + β6 RETURN t-2 + β7 VOLATILITY t-1 + β8 VOLATILITY t-2 + β9 VOLUME t-1 + β10 VOLUME t-2 + (β11 DAY1t + … + β14 DAY4t) + β15 MESSAGES t-1 + β16 LENGTH t-1 + β17 SENTI t-1 + β18 VARSENTI t-1 + β19 SUBJ t-1 + β20 VARSUBJ t-1 + εt Where t = days (t = 1, 2, …, n); day of the week (d = 1, …, 4); stakeholder clusters (s = 1, 2, …, c) 31 Experimental Design: Description of Prediction Models Stakeholder models Comprehensive baseline variables plus stakeholder grouplevel forum measures RETURN t = β0 + β1 FFSIZE + β2 FFBTM + β3 FFMARKET t-1 + β4 FFMARKET t-2 + β5 RETURN t-1 + β6 RETURN t-2 + β7 VOLATILITY t-1 + β8 VOLATILITY t-2 + β9 VOLUME t-1 + β10 VOLUME t-2 + (β11 DAY1t + … + β14 DAY4t) + (β15 MESSAGES 1 t-1 + β16 LENGTH 1 t-1 + β17 SENTI 1 t-1 + β18 VARSENTI 1 t-1 + β19 SUBJ 1 t-1 + β20 VARSUBJ 1 t-1 + … + βk MESSAGES c t-1 + βk+1 LENGTH c t-1 + β k+2 SENTI c t-1 + β k+3 VARSENTI c t-1 + β k+4 SUBJ c t-1 + β k+5 VARSUBJ c t-1) + εt Where t = days (t = 1, 2, …, n); day of the week (d = 1, …, 4); stakeholder clusters (s = 1, 2, …, c); index k = (((c - 1) * 6) + 15) 32 Experimental Design: Social Media Data A 17 month period was utilized for analysis and experimentation November 1, 2005 to March 31, 2007 First five months were utilized to calibrate the initial stock return prediction models November1, 2005 – March 31, 2006 Calibrated models applied for prediction during each trading day in the next month Each subsequent month, new models were calibrated using five previous months of time-series variables, for stock return prediction during the next month of trading In total, stock return prediction was performed daily for one year (250 trading days) April 1, 2006 – March 31, 2007 Forum Yahoo Finance – WMT (finance.yahoo.com) Wal-Mart Blows (www.walmartblows.com) Wakeup Wal-Mart (www.wakeupwalmart.com) Messages Discussion Threads Stakeholders Messages per Thread Messages per Stakeholder 134,201 40,633 5,533 3.30 24.25 55,125 3,690 1,461 14.94 37.73 10,797 1,306 915 8.27 11.80 33 Results and Discussion Hypothesis testing results Hypothesis Result H1.1 Baseline-Comp model > Baseline-FF model Partially supported H1.2 Baseline-Comp model > Baseline-Tech model Rejected H2 Forum-level models > best baseline models Rejected H3.1 Stakeholder-level models > best baseline models Supported H3.2 Stakeholder-level models > forum-level models Partially supported H4.1 Social network > discussion content representation Partially supported H4.2 Writing style > discussion content representation Rejected H4.3 Social network > writing style representation Partially supported H5.1 ANN > OLS Rejected H5.2 SVR > OLS Partially supported H5.3 SVR > ANN Partially supported 34 Results and Discussion Wal-Mart stock return prediction model results Baseline models using fundamental and technical variables Results across 250 trading days forecasted Baselines for simulated trading (initial investment of $10,000): Holding Wal-Mart stock for the year results in $10,096 Holding S&P 500 for the year results in $11,012 Model Baseline-FF Baseline-Tech Baseline-Comp OLS $ $ 9,787 $ 8,799 $ 10,763 OLS Accuracy 55.20% 57.20% 54.40% ANN $ $ 9,998 $ 9,702 $ 10,418 ANN Accuracy 44.40% 57.60% 56.80% SVR $ $ 9,408 $ 9,503 $ 10,645 SVR Accuracy 51.20% 56.40% 56.80% 35 Results and Discussion Wal-Mart stock return prediction model results Incorporating the Wakeup Wal-Mart web forum Results across 250 trading days forecasted Model Best Baseline Forum Stakeholder-SN Stakeholder -Content Stakeholder -Style Stakeholder-SN+Content Stakeholder-SN+Style Stakeholder-Content+Style Stakeholder-SN+Content+Style OLS $ $ 10,763 $ 10,367 $ 9,873 $ 10,689 $ 10,271 $ 10,384 $ 10,744 $ 10,696 $ 10,976 OLS Accuracy 57.20% 57.60% 55.20% 60.40% 56.00% 61.60% 60.00% 59.20% 58.00% ANN $ $ 10,418 $ 10,397 $ 10,930 $ 11,595 $ 9,653 $ 13,066 $ 10,792 $ 10,590 $ 10,778 ANN Accuracy 57.60% 59.20% 57.20% 60.40% 56.80% 60.80% 60.40% 56.40% 56.40% SVR $ $ 10,645 $ 10,303 $ 10,669 $ 11,976 $ 9,305 $ 11,866 $ 11,249 $ 10,603 $ 10,881 SVR Accuracy 56.80% 59.20% 59.20% 61.20% * 56.00% 62.80% ** 57.60% 58.80% 59.60% Pair-wise t-test; improvement over best baseline model at * p < 0.10 ** p < 0.05 36 AZ STOCK TRACKER III Introduction Forward-looking statements (FLS) refer to Projections, forecasts, or other predictive statements Made by firm management Section 21E of the Securities Exchange Act (1934) Extended forward-looking statements (EFLS) Statements that may have implications for a firms future development Similar to FLS, but broader Including information from information intermediaries (e.g., newspapers, newswires) and individuals (e.g., blogs) 38 Recognizing EFLS EFLS: Extends FLS to include statements about firm’s future performance from other sources such as financial press, analysts’ reports, and individuals Goal Recognition Task Definition EFLS Recognition Future Timing (FT) Primary content is about future events or states Explicit Uncertainty (EU) Explicit accounts of doubt or unreliability Overall Assessment (ALL) Affect decision maker’s belief about a firm’s future cash flow Positive (POS) Positive impact on the belief Negative (NEG) Negative impact on the belief EFLS Sentiment 39 AZ STOCK TRACKER III: EFLS 40 Summary of Annotation Results Agreement ALL 0.91 (0.88, 0.93) POS 0.90 (0.88, 0.93) NEG 0.89 (0.86, 0.91) • High kappa values (>0.7) on risks supports the coding 0.81 scheme being empirically (0.76, 0.86) valid 0.79 (0.73, 0.85)• Agreement upper bound • 89% to 91% (for ALL, 0.77 (0.71, 0.82) POS, and NEG) Cohen’s Kappa Category Count Percent ALL 1157 46% POS 836 33% NEG 904 36% Note: (95% CI) from 1,000 Bootstrappings • Reference Standard Dataset: – 2539 sentences in total 41 Experiment 1: Sentence-Level Evaluation Model Accuracy† F-Measure‡ Recall‡ Precision‡ LASSO 67.1% 66.5% 83.8% 55.1% ENET75 69.3% 68.0% 87.7% 55.6% ENET50 68.9% 68.7% 90.5% 55.4% ENET25 69.4% 68.9% 91.2% 55.4% SVM 69.5% 70.2% 83.9% 60.3% SVM w/IG 69.1% 68.9% 84.3% 58.3% FKC 64.7% 50.9% 69.7% 40.1% OF_PN 54.8% 27.9% 19.1% 51.4% 42 EFLS Impacts: Hypotheses Development Theoretical framework (Easley and O’Hara, 2004) There are 𝐼𝑘 signals for stock k (𝑠𝑘1 , 𝑠𝑘2 , … , 𝑠𝑘𝐼 ) 𝑘 1 𝑣𝑘 , 𝛾𝑘 𝑠𝑘𝑖 ~𝑁 (𝑠𝑘1 , 𝑠𝑘2 , 𝑠𝑘3 , 𝑠𝑘(𝛼𝑘𝐼𝑘) , 𝑠𝑘(𝛼𝑘 𝐼𝑘+1) , … , 𝑠𝑘(𝐼𝑘 −1) , 𝑠𝑘𝐼𝑘 ) Private Signals Public Signals 𝛼𝑘 : The relative amount of private-versus-public information 43 Hypotheses Development (Cont’d.) Hypothesis 1: Firms with lower EFLS intensity are associated with higher expected return. 𝜕𝐸[𝑣𝑘 − 𝑝𝑘 ] 𝛿𝑥𝑘 1 − 𝜇𝑘 𝐼𝑘 𝛾𝑘 = 2 𝜕𝛼𝑘 𝐶𝑘 1 + 𝛼𝑘 𝐼𝑘 𝜂𝑘 𝜇𝑘2 𝛾𝑘 𝜎 −2 2 >0 44 Hypotheses Development (Cont’d.) Hypothesis 2: Firms with lower EFLS intensity are associated with the higher stock volatility. 𝜕𝑉𝑎𝑟(𝑣𝑘 − 𝑝𝑘 ) 𝛿 4 𝛾𝑘 𝐼𝑘 1 − 𝜇𝑘 2𝛿 4 + 𝑉1,𝑘 + 𝑉2,𝑘 = 𝜕𝛼𝑘 𝜂𝑘 𝛿 2 𝜌𝑘 + 𝛾𝑘 𝐼𝑘 (1 + 𝛼𝑘 (𝜇𝑘 − 1)) + 𝛼𝑘 𝜂𝑘 𝛾𝑘 𝐼𝑘 𝜇𝑘2 (𝛾𝑘 𝐼𝑘 + 𝜌𝑘 ) 𝑉1,𝑘 = 𝛾𝑘 𝐼𝑘 − 𝜌𝑘 + 𝜇𝑘 𝛾𝑘 𝐼𝑘 + 𝜌𝑘 3 𝛼𝑘 𝜂𝑘2 𝐼𝑘 𝛾𝑘 𝜇𝑘2 + 𝛿 2 𝜂𝑘 𝑉2,𝑘 = −1 + 2𝜇𝑘 + 𝜇𝑘2 𝛿 2 𝜂𝑘 𝛾𝑘 𝐼𝑘 𝛼𝑘 𝜕𝑉𝑎𝑟 𝑣−𝑝𝑘 𝜕𝛼𝑘 If 𝐼𝑘 𝛾𝑘 > 𝜌𝑘 and 𝜇𝑘 > 2 − 1 then Intuition: if there are enough signals and the fraction of informed investors is larger than 41%, then firms with lower amounts of EFLS Higher Volatility >0 45 Control Variables Variable Definition Number of news articles mentioning firm i in month t. Logarithm of market value, computed using the closing market price of month t-1. Logarithm of book-to-market ratio, computed following Fama and French (1993). Log(Dollar trading volume of firm i in month t) Log(variance); variance of firm i in month t is computed using daily stock returns. Proportion of individual ownership of stock i, using the latest available data, computed by aggregating 13f filings (Fang and Peress 2009). Log(1+number of analysts covering firm i in month t). Log(1+standard deviation of analyst’s earnings predictions). Firm-Level Performance Evaluation (Cont’d.) Empirical Model 1: Hypothesis 1 Predicts Negative b1 𝑟𝑖,𝑡+1 = 𝑎0 + b1 𝐴𝐿𝐿_𝐼𝑁𝑖,𝑡 + 𝑐1 𝑁𝑒𝑤𝑠𝐹𝑟𝑒𝑞𝑖,𝑡 + 𝑐2 𝑁𝑒𝑤𝑠𝑆𝑒𝑛𝑡𝑖𝑖,𝑡 + 𝑑1 𝐿𝑜𝑔𝑆𝑖𝑧𝑒𝑖,𝑡 + 𝑑2 𝐿𝑜𝑔𝐵𝑀𝑖,𝑡 + 𝑑3 𝑟𝑖,𝑡 + 𝑑4 𝐿𝑜𝑔𝑉𝑖,𝑡 + 𝑒𝑖𝑡 Empirical Model 2: Hypothesis 2 Predicts b1 ≠ 0 𝐿𝑜𝑔𝑉𝑖,𝑡+1 = 𝑎0 + b1 ALL_INi,t + 𝑐1 𝑁𝑒𝑤𝑠𝐹𝑟𝑒𝑞𝑖,𝑡 + 𝑐2 𝑁𝑒𝑤𝑠𝑆𝑒𝑛𝑡𝑖𝑖,𝑡 + 𝑑1 𝐿𝑜𝑔𝑉𝑜𝑙𝑢𝑚𝑒𝑖,𝑡 + 𝑑2 𝐿𝑜𝑔𝑉𝑖,𝑡 + 𝑑3 𝐿𝑜𝑔𝑆𝑖𝑧𝑒𝑖,𝑡 + 𝑑4 𝐿𝑜𝑔𝐵𝑀𝑖,𝑡 + 𝑑5 𝑟i,t + 𝑑6 𝐼𝑛𝑑𝑣𝑂𝑤𝑛𝑖,𝑡 + 𝑑7 𝐿𝑜𝑔𝐴𝑛𝑎𝑙𝑦𝐶𝑜𝑣𝑒𝑟𝑖,𝑡 + 𝑑8 𝐿𝑜𝑔𝐴𝑛𝑎𝑙𝑦𝑆𝐷𝑖,𝑡 + 𝑒𝑖,𝑡 47 Experiment Two: Firm-Level Evaluation Research Testbed: January 1986 to May 2008, 1,134,321 Wall Street Journal news articles Merged with CRSP, Compustat, and IBES Stock prices lower than $5 at the end of a month were removed (Cohen and Frazzini 2008; Fang and Peress 2009) 1,274,711 firm-months, spanning 269 months 48 Expected Return and EFLS Intensity Variable Value -0.0026* Variable Value -0.0052** Variable Value -0.0039 Control Variables Intercept ***, **, * 0.00069*** 0.00068*** 0.00067*** -0.00081 -0.0012 -0.0015 -0.0019** -0.0019*** -0.0019*** 0.0025*** 0.0025*** 0.0025*** -0.046*** -0.046*** -0.046*** 0.00042 0.00042 0.00042 0.039*** Intercept 0.039*** Intercept 0.039*** 0.0031 0.0031 0.0031 indicate statistical significance at the 0.01, 0.05, and 0.1 levels, respectively. 49 Volatility and EFLS Intensity Model 2A (𝐴𝐿𝐿_𝐼𝑁𝑖,𝑡 ) Variable 𝐴𝐿𝐿_𝐼𝑁𝑖,𝑡 Value -0.074*** Model 2B (𝐹𝑇_𝐼𝑁𝑖,𝑡 ) Variable 𝐹𝑇_𝐼𝑁𝑖,𝑡 Model 2C (EU_𝐼𝑁𝑖,𝑡 ) Value -0.196*** Variable 𝐸𝑈_𝐼𝑁𝑖,𝑡 Value -0.254*** Control Variables 𝑁𝑒𝑤𝑠𝐹𝑟𝑒𝑞𝑖,𝑡 𝑁𝑒𝑤𝑠𝑆𝑒𝑛𝑡𝑖𝑖,𝑡 𝐿𝑜𝑔𝑉𝑜𝑙𝑢𝑚𝑒𝑖,𝑡 𝐿𝑜𝑔𝑉𝑖,𝑡 𝐿𝑜𝑔𝑆𝑖𝑧𝑒𝑖,𝑡 𝐿𝑜𝑔𝐵𝑀𝑖,𝑡 𝑟𝑖,𝑡 𝐼𝑛𝑑𝑣𝑂𝑤𝑛𝑖,𝑡 𝐿𝑜𝑔𝐴𝑛𝑎𝑙𝑦𝐶𝑜𝑣𝑒𝑟𝑖,𝑡 𝐿𝑜𝑔𝐴𝑛𝑎𝑙𝑦𝑆𝐷𝑖,𝑡 Intercept 𝑅2 ***, **, * 0.012*** -0.105*** 0.108*** 0.565*** -0.222*** -0.066*** -0.615*** 0.071*** 0.016*** 0.095*** -1.568*** 0.57 𝑁𝑒𝑤𝑠𝐹𝑟𝑒𝑞𝑖,𝑡 𝑁𝑒𝑤𝑠𝑆𝑒𝑛𝑡𝑖𝑖,𝑡 𝐿𝑜𝑔𝑉𝑜𝑙𝑢𝑚𝑒𝑖,𝑡 𝐿𝑜𝑔𝑉𝑖,𝑡 𝐿𝑜𝑔𝑆𝑖𝑧𝑒𝑖,𝑡 𝐿𝑜𝑔𝐵𝑀𝑖,𝑡 𝑟𝑖,𝑡 𝐼𝑛𝑑𝑣𝑂𝑤𝑛𝑖,𝑡 𝐿𝑜𝑔𝐴𝑛𝑎𝑙𝑦𝐶𝑜𝑣𝑒𝑟𝑖,𝑡 𝐿𝑜𝑔𝐴𝑛𝑎𝑙𝑦𝑆𝐷𝑖,𝑡 Intercept 𝑅2 0.012*** -0.103*** 0.108*** 0.565*** -0.222*** -0.066*** -0.615*** 0.071*** 0.017*** 0.095*** -1.566*** 0.57 𝑁𝑒𝑤𝑠𝐹𝑟𝑒𝑞𝑖,𝑡 𝑁𝑒𝑤𝑠𝑆𝑒𝑛𝑡𝑖𝑖,𝑡 𝐿𝑜𝑔𝑉𝑜𝑙𝑢𝑚𝑒𝑖,𝑡 𝐿𝑜𝑔𝑉𝑖,𝑡 𝐿𝑜𝑔𝑆𝑖𝑧𝑒𝑖,𝑡 𝐿𝑜𝑔𝐵𝑀𝑖,𝑡 𝑟𝑖,𝑡 𝐼𝑛𝑑𝑣𝑂𝑤𝑛𝑖,𝑡 𝐿𝑜𝑔𝐴𝑛𝑎𝑙𝑦𝐶𝑜𝑣𝑒𝑟𝑖,𝑡 𝐿𝑜𝑔𝐴𝑛𝑎𝑙𝑦𝑆𝐷𝑖,𝑡 Intercept 𝑅2 indicate statistical significance at the 0.01, 0.05, and 0.1 levels, respectively. 0.012*** -0.110*** 0.108*** 0.565*** -0.222*** -0.066*** -0.616*** 0.071*** 0.017*** 0.095*** -1.566*** 0.57 50 Take-Away and WIP (20%) Mass and social media texts provide additional signals for market prediction (in addition to numbers) Message volume important; aggregate sentiment may not (EMH) Business sentiment processing difficult; may require additional content pre-processing (stakeholder; EFLS) Predicting return hard; predicting volatility easier (VIX Chicago Board) Large-scale stock news tracking and text analytics can be automated Trading windows; decay function; targeted sentiment; extensive trading periods (up/down); industry and news category (oil/banking); firm & index size (Russell/NYSE); emerging markets (China) All the firms (10K), all the news (1M each), all the time ??? Trading strategy ??? 51 SEC/Edgar NYSE.com NASDAQ.com Finance.Yahoo.com Company Information Database Ticker CIK CUSIP Company Name PERMNO Yahoo Finance Forums Company Websites Twitter Stock Exchange WSJ Dynamic Data Sources Search Engines 10K Report Blogs News Data Processing Transformation/Integration Performance Indicators Topics & Sentiments Time Series / Burst Risk Model SNA Data Analysis Interactive Applications Data Collection Predefined Data Sources Company Keywords Static Figures/Dashboards Basic Information Data Sources for US Public Companies Analytic Approaches Single Media Analysis Cross Media Analysis Predictive Analysis Simulated Trading 52 AZ BIZ INTEL System Design Visualization Hsinchun Chen, Ph.D. Artificial Intelligence Lab, University of Arizona hchen@eller.arizona.edu http://ai.arizona.edu