The Influence of News Articles on The Stock Market.

Transcription

1 The Influence of News Articles on The Stock Market. COMP4560 Presentation Supervisor: Dr Timothy Graham U Zhiheng Zhou Australian National University At Ian Ross Design Studio On

2 Motivation Stock market prediction is a hot topic. News may influence stock market?? Influence 2

3 Project background Introduction: Work on news title and stock market data. Sentiment function. Machine learning. Justify the news influence on stock market. 3

4 Problems state: Do fluctuations in news sentiment (news sentiment series ) influence fluctuations in the stock market (stock market series)? Sentiment part Can a dictionary-based sentiment method be improved using a thesaurus and grammatical features? Sentiment part Can my machine learning model have a good performance on stock market? Machine learning part 4

5 Background - Data set News of Apple from to (about 2 years) Stock market data of Apple. (opening price, closing price, volume) 5

6 Background Previous Research [1] Use neural network, two class ( +1, -1 ) Stock market data: S&P500 [1] Xiao Ding, Yue Zhang, Ting Liu, and Junwen Duan. Using structured events to predict stock price movement: An empirical investigation. In Proc. of EMNLP, pages , Doha, Qatar, October Association for Computational Linguistics. 6

7 Method Sentiment part Original Sentiment Function: Hu and Liu dictionary [1]. Has 2006 positive words and 4783 negative words. Bag of words. Remove stop-words, punctuation. Match words to the dictionary. Positive, score + 1, negative, score -1. Baseline [1] Hu, M., & Liu, B. (2004). Mining and summarizing customer reviews. Paper presented at the KDD2004 conference, pp

8 Limitations of Original Sentiment Function: Dictionary size not big enough - The total size of Hu and Liu dictionary (around 6800 words). Only match single word but not phrase Some phrases may have positive or negative state. (e.g. look forward to) Remove stop-words and punctuation so that Cannot detect the grammar state some stop-words may change the sentiment state. (e.g. not) 8

9 A modified sentiment function using Thesaurus + not Grammar modifier My Sentiment Function: Base on original sentiment function using Hu and Liu dictionary. Extended Hu and Liu dictionary by the thesaurus dictionary. Using original news text, not removing stop-words and punctuation. Using not grammar modifier. 9

10 Evaluation - Granger causality test [1]: An evaluation technique on two series to test the correlations between them. Using on score series and stock market series. News may take a while before it influence stock market - Using lags from 1 to 10 [1] 10

11 Granger Causality test and Lag example: Source: By BiObserver - Own work, CC BY-SA 3.0, 11

12 Implementation Sentiment Using Thesaurus dictionary to extend original dictionary: 12

13 not grammar modifier: not will opposite the sentiment state for the following words. (e.g. not happy ) Effect Scope : From itself to the next comma or full stop. Using a Tag to multiply the sentiment score of a word. In effect scope: Tag = -1, otherwise, Tag = 1 13

14 Method Machine learning part A supervised Machine Learning approach. Attributes : News term matrix. Label : Stock market data movement. Model method: xgbtree extreme Gradient boosting xgboost. K-fold cross-validation Source: 14

15 Challenges of Time Series data: Times order matters - Cannot obtain future data from past. Normal k-folds cross validation split data randomly - break data s natural order. Solution: Training set - past data. Testing set - future data. 15

16 K-Folds Cross Validation for Time Series Data: Past Times Future 1/2 K = 2 Training 2/3 Testing K = 3 Training dataset Training Testing 9/10 K = 10 Training Testing 16

17 Feature extraction Encode continuous data to category data: Motivation - Machine learning is more capable for classify problem Delta of data - Today s data minus yesterday s data. Encode the Delta as some movement trend classes Delta >> 0, trend class up Delta << 0, trend class down Delta close to 0, trend class keep 17

18 Implementation Machine learning Training parameters: Using time-slice for cross validation. (A cross validation method for time series) Using default parameters 18

19 Experiment setup: Language : R IDE : R studio OS : macos CPU : 2.8 GHz Intel Core i7 RAM : 16G 19

20 Results Sentiment Granger Causality test results: Null Hypothesis X not Granger cause Y. (X is sentiment score, Y is opening price, closing price and volume) P-value the probability of Null Hypothesis Source: 20

21 Volume Results: Methods Lags P-value(Original) P-value(My method) ** ** * * * * Over all, my method better than original. For my method, Lag = 1, p-value smallest. 21

22 Discussion: Both sentiment score (original and my) do not Granger cause opening price series and closing price series. Sentiment score series of my sentiment method can Granger cause volume series with high probability. Opening price and closing price my influenced by some deeper factors like policy and economic. Volume directly influenced by people, people can directly influenced News. 22

23 Limitations: Need a great amount experiment on others dataset. Granger Causality justify the statistic relationship between two series, but not the real relationship. Improvements: Add more grammar modifier. Use n-grams. Use other method to extend dictionary. Use more advanced method for sentiment. Use advanced technique to test on sentiment score. 23

24 Motivation of Machine learning part: Sentiment method is not enough. Seek more relationship of news and stock market Brief machine learning technique on data. 24

25 Results Machine Learning: Three class up, keep, down Using volume as label Lag = 1 Overall Accuracy Precision up 0.25 Precision keep 0.60 Precision down 0.16 Recall up Recall keep Recall down

26 Discussion: Overall not good results. Class keep performance better than others two. The number of class keep much more than others. Model tend to class most data to keep Limitations: Feature extraction method is manually. Just use lag 2 on volume Use default parameters. Improvements: Tuning parameters Use advanced method on feature extraction. 26

27 Recall Problems state: Do fluctuations in news sentiment (news sentiment series ) influence fluctuations in the stock market (stock market series)? Sentiment part Can a dictionary-based sentiment method be improved using a thesaurus and grammatical features? Sentiment part Can my machine learning model have a good performance on stock market? Machine learning part 27

28 Conclusions: Fluctuations in news sentiment has a great probability influence (Granger Cause) fluctuations in the stock market volume. For opening price and closing price, more research. My sentiment method has a great chance to improve original dictionary-based sentiment method. My model performance on volume is not good. 28

29 Future works: Use NLP as sentiment part Use more method on machine learning part. Use the sentiment score as an attributes in machine learning part. Use neural network. 29

30 Thank you. Acknowledgements: Dr Timothy Graham Professor Peter Strazdins