Review on financial document sentiments

Slides:



Advertisements
Similar presentations
Stock Price Prediction Based on Social Network A survey Presented by: CHEN En.
Advertisements

Prediction Modeling for Personalization & Recommender Systems Bamshad Mobasher DePaul University Bamshad Mobasher DePaul University.
1/19 Motivation Framework Data Regressions Portfolio Sorts Conclusion Option Returns and Individual Stock Volatility Jie Cao, Chinese University of Hong.
Supervised Learning Techniques over Twitter Data Kleisarchaki Sofia.
Fundamental Analysis Workshop Series Session Five – Dividend Investing.
Joey Engelberg University of California - San Diego Financial Risks International Forum March 21, 2014 Search Data and Behavioral Finance.
1 Module 7 Fundamental Analysis. 2 Module 7 - Learning Objectives Define fundamental analysis. Differentiate between fundamental, technical and speculative.
Introduction to Modern Investment Theory (Chapter 1) Purpose of the Course Evolution of Modern Portfolio Theory Efficient Frontier Single Index Model Capital.
Twitter Mood Predicts the Stock Market Authors: Johan Bollen, Huina Mao, Xiao-Jun Zeng Presented By: Krishna Aswani Computing ID: ka5am.
Stock Market Game Current Events.
Twitter Volume Spikes: Analysis and Application in Stock Trading Yuexin Mao, Wei Wei and Bing Wang COMP4332/RMBI4310 CHAN Chun Ting ( )
Mr. Lange - Economics.  Welcome to Mr. Lange’s Stock Market Simulation!  As members of an investment group, you will be competing against one another.
Forecasting with Twitter data Presented by : Thusitha Chandrapala MARTA ARIAS, ARGIMIRO ARRATIA, and RAMON XURIGUERA.
Stock Valuation – Technical Analysis Essentials of Corporate Finance Chapters 7 and 10 Materials Created by Glenn Snyder – San Francisco State University.
Identifying Good Stock Investments Investment and Finance 12 Ms. Stewart.
Stock Market Prediction Using Sentiment Detection C. LEE FANZILLI ADVISORS: PROF. DVORAK AND PROF. WEBB.
More than words: Social networks’ text mining for consumer brand sentiments A Case on Text Mining Key words: Sentiment analysis, SNS Mining Opinion Mining,
Types of Investments Stocks Bonds Mutual Funds Real Estate Savings/Certificates of Deposit Collectibles.
Cross-Cultural factors and Portfolio Choice Daniel Egan, Greg Davies, Peter Brooks Barclays Wealth Behavioural Analytics FUR Conference 01/07/2008.
Designing Ranking Systems for Consumer Reviews: The Economic Impact of Customer Sentiment in Electronic Markets Anindya Ghose Panagiotis Ipeirotis Stern.
Prediction of Influencers from Word Use Chan Shing Hei.
Free boundary value problems in mathematical finance presented by Yue Kuen Kwok Department of Mathematics December 6, 2002 * Joint work with Min Dai and.
Market Efficiency.
CISC 849 : Applications in Fintech Financial Visualization Leonardo De La Rosa Institute for Financial Services Analytics University of Delaware.
McGraw-Hill/Irwin Copyright © 2013 by The McGraw-Hill Companies, Inc. All rights reserved. CHAPTER PLAYLIST SONG: “BIG MONEY“BIG MONEY” BY GARTH BROOKS.
Content Marketing Proposal NAME AND CLASS. Outline Part I: Market ◦The Company ◦Product/Service ◦Current Online Status ◦Challenges ◦Opportunity Part II:
PREDICTING STOCK MARKET MOVEMENT USING SENTIMENTS For EECSE 6898-From Data to Solutions class Presented by-Tulika Bhatt(tb2658)
More than words: Social network’s text mining for consumer brand sentiments Expert Systems with Applications 40 (2013) 4241–4251 Mohamed M. Mostafa Reporter.
Security Valuation & Investment Decision With Zaveri Securities Ltd. Sagar V Moradiya Roll no-43 1 Guide : Prof. Sandhya Harkawat.
Twitter Based Research Benny Bornfeld Mentors Professor Sheizaf Rafaeli Dr. Daphne Raban.
Prepared by Fayes Salma.  Introduction: Financial Tasks  Data Mining process  Methods in Financial Data mining o Neural Network o Decision Tree  Trading.
Early Warning and Real Time Anomaly Detection System © 2015 Blueocean Market Intelligence1.
BUS 308 Week 5 Final Paper Check this A+ tutorial guideline at 308-Week-5-Final-Paper Writing the Final Paper.
Valuation: Market-Based Approach
Presenter: Siddharth Krishna Sinha Instructor: Jing Gao
Detection of Misinformation on Online Social Networking
Does Academic Research Destroy Stock Return Predictability. R
An Analysis of Critical Accounting Policies
Tactics II – Volatility & Time Iron Condors
CISI – Financial Products, Markets & Services
The Sellout: Readers Sentiment Analysis of 2016 Man Booker Prize Winner Paper ID : 748.
INTRODUCTION TO EQUITIES
Market Intelligence Analysis
Stock Market 101 What are stocks? What is the stock market?
Contextual Intelligence as a Driver of Services Innovation
Sentiment analysis tools
Textural sentiment in finance
OPTIMAL SHOUTING POLICIES OF OPTIONS
Predictability Using Big Data
MID-SEM REVIEW.
Correlating Stock Price Shifts with Predictions from Twitter
Forecasting Exchange Rates
Financial Algebra 25 April 2018.
Event Studies.
Bell Ringer List as many things that come to mind when you hear the words “stock market” or “stocks.”
المركز الشرقي للاستشارات Orient Consulting Center مؤسسة استشارية كويتية تأسست عام 1984، ومنذ ذلك التاريخ قدمت العديد من الخدمات لقائمة كبيرة من الشركات.
Introduction to Modern Investment Theory (Chapter 1)
Financing and Investing
Text Analysis and Search Analytics
Economics 434: The Theory of Financial Markets
How often will the Monkey’s portfolio beat the market average?
The information Content of IPO Prospectuses
Part III Exchange Rate Risk Management
Statistical n-gram David ling.
What do they mean and how can I use them?
Megaputer Intelligence
Text Analysis and Search Analytics
Applied Machine Learning For Quant Finance
Big Data Big Data first appeared towards the end of the 1990’s and has become a buzz word in the last few years.
RHO Index Key Levels 13 June 2019.
Presentation transcript:

Review on financial document sentiments 24-8-2017 David Ling HSMC Deep Learning Center

Contents Financial sentiment examples Brief history Performance Dictionary based Machine learning/ deep learning Our plan

What is financial document sentiment Financial document sentiment – getting tones or emotions from the document Annual reports, news, twitters, analyst reports, forum posts, etc. Predictions on stock prices and returns, financial distress, risks, and volatility (Feng Li, SSRN 2006; Gabriele Ranco, PLOS ONE 2015; Steven L., FFJ 2017; Shimon Kogan, NAACL 2009; Petr Hájek, EANN 2013; Y Liu, arXiv 2017) “匯控獲大摩唱好,公司回購股份力度加碼,再升3.3%,收65.5元,推動恒指升逾78點,即…” This is good! “《收市總結》港股連續兩日成交不足500億元 全周累跌172點…” This is bad!

Texts store views of the past and future L. Iliadis, H. Papadopoulos, and C. Jayne EANN 2013 Texts store views of the past and future Annual reports ->Extract numerical data and sentiment scores Investment grade (IG) and non-investment grade (NG) Assigned by a highly regarded Standard & Poor’s rating agency 520 US companies in 2010, predicting results in 2011 Negative: difficult Litigious: sue, sued, summon Uncertainty: may Weak Modal: could, may, probable Strong modal: must

First day return = closing price – offer price Tim Loughran, Journal of Financial Economics 2013 S-1 Filling - a form filled for the exchange before going public in US (~50000 words) Calculated the correlation between S-1 sentiment scores and first day returns First day return = closing price – offer price Includes 1,887 completed U.S. IPOs with an offer price of at least $5 per share during the 1997–2010 time period Weak modal and positive tones have relatively higher correlation => will have a higher return Weak modal: may, could, might, possible Correlations to IPO first day returns

News sentiment Reuters NewScope Sentiment Engine Steven L, Financial analysis Journal, 2017 Reuters NewScope Sentiment Engine Data service provided by Reuters (non-free) Real time news updating with sentiment scores 3 indices: Pos, Neut, Neg Tagged with date and related company http://share.thomsonreuters.com/assets/elektron/news-analytics-flyer.pdf

News sentiment Trade base on the news sentiment score A company score is obtained by averaging the related news score Long the top 20% and short the bottom 20% according to the scores (daily or weekly) x day after news means using scores x days before. x = 1 means using yesterday’s score 0.17% for 1 day after, 0.32% for 1 week after Positive return for x < 0 implies news stories may lag events that affect stock prices Daily news can predict 1 day return only, while weekly based shows a better result Daily return Weekly return

News sentiment 優礦 A mainland big data company Provide news sentiment scores of a company via API 40k news threats per day Sentiment score: [-1,1]

Gabriele PLOS 2015 Twitter Sentiment Analyse 15 months, 30 companies in DJIA, total 1.5M tweets from users (eg. McDonald’s, Visa, Coca-cola) The machine learning classifier tells you whether a tweet is negative, neutral, or positive Two time series, daily sentiment score 𝑃 𝑑 and daily stock return 𝑅 𝑑 Small non-zero linear correlation is obtained Trivago 3M Part of the correlations for different companies (Gabriele PLOS 2015)

Twitter sentiment Event study Events (eg. Earning announcement) are classified according to the number of positive and negative tweets returns are affected by events the significantly An even can be classified by using tweets sentiment Event day Result obtained by averaging across all companies and events.

Brief history on financial document sentiment Early financial document sentiment (from ~2006) Official documents: annual/quarter reports (10-Ks), analyst reports Newspaper: The Wall Street Journal, New York Times Dictionary based Feng Li, SSRN 2006; Henry, SSRN 2009; Tim Loughran, JFE 2013; Petr Hájek, EANN 2013; Li Xiaodong, 2014 Recent years Online news (Reuters) Social media (Tweeters, StockTwits, forum messages) Machine learning/ Deep learning Matthias, JBF 2014; Gabriele Ranco, PLOS ONE 2015; Steven L. Heston, FAJ 2017; Y Liu, arXiv 2017 Growing number of companies are implementing the technology, but mainly for English, and not for Hong Kong

Dictionary based Sentiment scores are calculated by Keyword frequency in the text Positive: excellent, nice, agree, etc. Negative: against, afraid, etc. Common keyword lists: Harvard IV-4 categories: positive, negative, strong, weak, active, passive Loughran & McDonald word lists (2011) 知网 (2007) Keyword usages are often different across disciplines and regions Eg. Taiwan news and Hong Kong news Eg. Financial reports and general text Subjective, not accurate, and not sophisticated

Dictionary based 台股新聞情緒指標 Ming Chuan University, Taiwan, 2013 Website format Two sentiment scores ITDC and SR Scores are obtained from Taiwan financial news Provide recommendation of companies Provide news sentiment testing “指標試算”

Dictionary based News sentiment 1: News sentiment 2: Incorrect, the tone should be negative. Correct, the tone is negative. Dictionary seems not so sophisticated. Like “蝕” is not detected.

Deep learning/ Machine learning Deep learning and machine learning Latest technology in extracting textual sentiment Don’t have to specify rules manually Able to understand semantic meaning More accurate, even better than human With more data, system becomes more sophisticated

Machine learning Machine learning classifier performance Gabriele PLOS 2015 Machine learning classifier performance Annotator agreement: comparing results of two human annotators Sentiment classifier: comparing results of the machine and a human Machine has a comparative accuracy, and a slightly higher 𝐹 1 score (excluding the neutral class) 𝐹 1 =

Deep learning Deep learning allows machine to learn semantic meaning Prepared 30000 pieces of online financial news from Quamnet (華富財經) Sample news: 長和今天放榜,早前大摩預測長和中期比只升5%,主要受英英鎊貶值 等外匯因素影響,預測長和上半年經營溢利同比升5%至309億元 Nearest to 跌: Nearest to 公布: 升, 挫, 倒跌, 微跌, 股亦收升, 現跌, 無升, 微升, 公佈, 宣布, 公告, 放榜, 發布, 止, 公在, Nearest to 同比: Nearest to 在: 按年, 去年同期, 之後高見, 僅減, 連特別息, 遠洋報, 此負, 於, 或, , 預期, 將在, 將於未來, 資源予, 與, Nearest to 而: Nearest to 對: 但, 另外, 表示, 或, 九鐵, 至於, 認為, 他稱, 認為, 家會員, 讓, 運費, 他們將, 將對, 令電能, 與, Nearest to 涉及: 共, 成交, xhand, 涉資約, 光啟, bcm_energy_partners, 對換,

Deep learning Better performance than simply machine learning method Published by google in 2014 Results on classifying movie reviews: 12.2%, much lower than the other Movie review sentiment comparison Quoc Le, Tomas Mikolov 2014

Our plan and project To implementing deep learning technologies to financial industry Providing sentiment scores of Hong Kong’s Chinese news More accurate and objective Handle huge amount of daily news and reports Recommendation Perfect right time (have not seen in Hong Kong market) Big market (we believe mainland analysts will also be interested)

Thank you Selected references: [Matthias 2014] Reuters Sentiment and Stock Returns, The Journal of Behavioral Finance, 15: 287–298 [Steven L. 2017] News vs. Sentiment: Predicting Stock Returns from News Stories, Financial Analysts Journal, 73 (3) [Tim Loughran 2013] IPO First-Day Returns, Offer Price Revisions, Volatility, and Form S-1 Language, Journal of Financial Economics (JFE), Forthcoming [Gabriele 2015] The Effects of Twitter Sentiment on Stock Price Returns, PLoS ONE 10(9) [Xiao 2015] Deep Learning for Event-Driven Stock Prediction, IJCAI [Petr Hájek 2013] Evaluating Sentiment in Annual Reports for Financial Distress Prediction Using Neural Networks and Support Vector Machines, EANN 2013, Part II, CCIS 384, pp. 1–10 [Quoc Le 2014] Distributed Representations of Sentences and Documents, arXiv:1405.4053 [cs.CL] [Jonas 2017] How the Market Can Detect Its Own Mispricing - A News Sentiment Index to Detect Irrational Exuberance, Proceedings of the 50th Hawaii International Conference on System Sciences [Shimom 2009] Predicting Risk from Financial Reports with Regression, NAACL