Embed Size (px)
Citation preview
Boosting Financial Trend Predictionwith Twitter Mood
Based on Selective Hidden Markov Models
Yifu Huang1, Shuigeng Zhou1∗, Kai Huang1 and Jihong Guan2
1Shanghai Key Lab of Intelligent Information ProcessingSchool of Computer Science, Fudan University{huangyifu, sgzhou, kaihuang14}@fudan.edu.cn
2Department of Computer Science and Technology, Tongji [email protected]
DASFAA 2015, Hanoi, Vietnam
Huang et al. (FDU CS) Boosting Financial Trend Prediction 04/23/2015 1 / 23
The Start
Accuracy: 91.967%
Huang et al. (FDU CS) Boosting Financial Trend Prediction 04/23/2015 2 / 23
The Start (Cont.)
But under certain circumstance
Huang et al. (FDU CS) Boosting Financial Trend Prediction 04/23/2015 3 / 23
Outline
1 Overview
2 Method
3 Experiment
4 Conclusion
Huang et al. (FDU CS) Boosting Financial Trend Prediction 04/23/2015 4 / 23
Overview
Summary
What?
Make more accurate and controllable stock prediction
Why?
Analyze and model the causality behind stock trendDesign and implement more practical prediction method
How?
Accuracy: exploit society moodControllability: adopt selective prediction
Huang et al. (FDU CS) Boosting Financial Trend Prediction 04/23/2015 5 / 23
Overview
Workflow
Massive Tweets
Mood Extraction via
POMS Bipolar and
WordNet
Twitter Moods
Mood Evaluation via
GCA
Financial
Growth Rates
Train Up
Multi-stream
sHMM
PredictFinancial Trends
Selected
Twitter Moods
Down
Don t
Know
Huang et al. (FDU CS) Boosting Financial Trend Prediction 04/23/2015 6 / 23
Method
Mood Extraction
Behavior finance
Individual mood -> individual decisionSociety mood -> society decision
Society mood measurement
Twitter, sense the world
POMS Bipolar Lexicon
Composed-anxious (Com.), agreeable-hostile (Agr.),elated-depressed (Ela.), confident-unsure (Con.),energetic-tired (Ene.), clearheaded-confused (Cle.)Expanding by WordNet synsets
Huang et al. (FDU CS) Boosting Financial Trend Prediction 04/23/2015 7 / 23
Method
Mood Extraction (Cont.)
Efficient extractation under MapReduce framework
Twitter data is large, so map it to different nodes, extract poms vectorfrom each tweet, and reduce them to overall poms index
Map (offset, line, date, poms_individual)
Filter
Ignore {http:, www.}, hold {i feel, makes me, ...}
Stem
Agreed -> agree, disabled -> disable, ...
Analyze
Seren -> composed, shaki -> anxious, ...
Reduce (date, poms_individual, date, poms_society)
Average
Huang et al. (FDU CS) Boosting Financial Trend Prediction 04/23/2015 8 / 23
Method
Mood Evaluation
Granger Causality Analysis
Determine whether one time series is useful in forecasting another
Y: growth rate of financial index; X: each Twitter mood
Yt = y0+lag
∑i=1
yiYt−i + εt (1)
Yt = y0+lag
∑i=1
yiYt−i +lag
∑i=1
xiXt−i + εt . (2)
Huang et al. (FDU CS) Boosting Financial Trend Prediction 04/23/2015 9 / 23
Method
Multi-stream sHMM
Hidden Markov Models -> HMMGenerative probabilistic model with latent states, where hidden statetransitions and visible observation emissions are assumed to be Markovprocesses
Selective prediction -> sHMMIdentify risk state set and prevent predictions that are made from them
Multiple stream -> Multi-stream sHMMTreat historical financial trend and Twitter mood trends as multipleobservation sequences generated by sHMM
Huang et al. (FDU CS) Boosting Financial Trend Prediction 04/23/2015 10 / 23
Method
Multi-stream sHMM - Training
Huang et al. (FDU CS) Boosting Financial Trend Prediction 04/23/2015 11 / 23
Method
Multi-stream sHMM - Training - Refine
Huang et al. (FDU CS) Boosting Financial Trend Prediction 04/23/2015 12 / 23
Method
Multi-stream sHMM - Prediction
Huang et al. (FDU CS) Boosting Financial Trend Prediction 04/23/2015 13 / 23
Method
Multi-stream sHMM (Cont.)
Large-scale performance evaluation
Random initialization number is large, so map Multi-stream sHMM todifferent nodes, get error rate from each model after train and predict,and reduce them to overall error rate
Map (offset, line, reject_bound, error_rate)
TrainPredict
Reduce (reject_bound, error_rate, reject_bound, avg_error_rate)
Average
Huang et al. (FDU CS) Boosting Financial Trend Prediction 04/23/2015 14 / 23
Experiment
Twitter Mood
Huang et al. (FDU CS) Boosting Financial Trend Prediction 04/23/2015 15 / 23
Experiment
Financial Index
Jun 15 Nov 09 Dec 21
0
Date
Gro
wth
Rat
e
S&P500NYSE
Jan 05 Mar 18 Jul 26
0
Date
Gro
wth
Rat
e
S&P500NYSE
Huang et al. (FDU CS) Boosting Financial Trend Prediction 04/23/2015 16 / 23
Experiment
Results of Granger Causality Analysis
Huang et al. (FDU CS) Boosting Financial Trend Prediction 04/23/2015 17 / 23
Experiment
Prediction Performance Comparison
Huang et al. (FDU CS) Boosting Financial Trend Prediction 04/23/2015 18 / 23
Experiment
Prediction Performance Comparison (Cont.)
Huang et al. (FDU CS) Boosting Financial Trend Prediction 04/23/2015 19 / 23
Experiment
Prediction Performance Comparison (Cont.)
Huang et al. (FDU CS) Boosting Financial Trend Prediction 04/23/2015 20 / 23
Conclusion
Conclusion and Future Work
Our method not only performs better than the state-of-the-artmethods, but also provides a controllability mechanism to financialtrend predictionExplore multivariate GCA to select the optimal combination ofmultiple Twitter moods to improve prediction performance
Huang et al. (FDU CS) Boosting Financial Trend Prediction 04/23/2015 21 / 23
Conclusion
The End
Thank you!
Huang et al. (FDU CS) Boosting Financial Trend Prediction 04/23/2015 22 / 23
Appendix
References I
[1] Dmitry Pidan, Ran El-Yaniv: Selective Prediction of FinancialTrends with Hidden Markov Models. NIPS 2011:855-863
[2] Johan Bollen, Huina Mao, Xiao-Jun Zeng: Twitter mood predictsthe stock market. J. Comput. Science (JOCS) 2(1):1-8 (2011)
[3] Jianfeng Si, Arjun Mukherjee, Bing Liu, Qing Li, Huayi Li, XiaotieDeng: Exploiting Topic based Twitter Sentiment for Stock Prediction.ACL 2013:24-29
Huang et al. (FDU CS) Boosting Financial Trend Prediction 04/23/2015 23 / 23
