tatiana silinskaya , ruslan novikov , evgeny polyakov , alexander porshnev, vladimir rossohin
DESCRIPTION
iCare’14, September, 21-24, Perm. Relationship between of emotional states of Twitter users and stock market indicators: search for causality. Tatiana Silinskaya , Ruslan Novikov , Evgeny Polyakov , Alexander Porshnev, Vladimir Rossohin - PowerPoint PPT PresentationTRANSCRIPT
Relationship between of emotional states of Twitter users and stock
market indicators: search for causality
Tatiana Silinskaya, Ruslan Novikov, Evgeny Polyakov, Alexander Porshnev, Vladimir Rossohin
National Research University Higher School of Economics, Nizhny Novgorod
iCare’14, September, 21-24, Perm
Twitter as data source Reaction on Associated Press tweet on White House explosions (Tue Apr 23, 2013 7:01pm EDT) http://www.reuters.com/article/2013/04/23/net-us-usa-whitehouse-ap-idUSBRE93M12Y20130423
https://blog.twitter.com/2012/a-new-barometer-for-the-election
Twitter Political Index: A Comparison to Gallup(with 30-day moving averages – August 1, 2010 – July 31, 2012
---- Twitter---- Gallup
Twitter - as moodmeter
Prediction System
Psycho-linguistics Math-linguistics (machine learning) Behavioral economics Econometrics
Sentiment analysis of twits
• Zhang et al. (2009) Correlation Hope+Fear+Worry%-3-mean DJIA = − 0.726** NASDAQ =− 0.728** S&P=− 0.713** VIX= 0.633** (98 days)
• Bollen et al. (2011) – 87.6%.for DJIA (0,1) (non filtered) weakness: only 21 days of testing
• Vu et al. (2012) – 82.93% for Apple (AAPL), Amazon (AMZN) 75.00%, Google (GOOG)- 80.49%, Microsoft (MSFT)- 75.61% (filtered by NER) weakness: only 41 days of testing
• Chen et al. (2013) “happy” and “sad” sentiment 70% accuracy Weakness: 33 days simulation
• Porshnev A. et al. (2013) – 68.60% (S&P500), 73.3% (DJIA) (from 75 up 256 days for testing) weakness no econometric model
Data (English language)• Period:
February 13, 2013 till September 29, 2013
• Finance historical data:
finance.yahoo.com
InversotPoint.com Historical data
• Twitter API
755’000 101 messages
MethodSimple sentiment analysis• Frequency of tweets with words from created emotional
dictionaries (8 emotions, 217 words). For example calm dictionary: impassive, motionless, stationary, still, calm, etc.
• Frequency of emotional words : worry, fear, hope etc.• Frequency of emoticons, interjections (lol, wth, etc.)
DJIA, S&P500, NASDAQ• Historical volatility• Trading volumes• Open Prices• Close Prices
Granger causality test (with lag from 1 to 7 days)
DJIA volatility
NASDAQ volatility
S&P500 volatility
Correlation between volatility and emotions
worry hope calm angry tired sad DJIA 0,41 0,29 0,13 0,36 0,37 0,31 NASDAQ 0,43 0,37 0,23 0,41 0,44 0,34 S&P500 0,45 0,35 0,20 0,40 0,41 0,33
Liquidity (trade volume)
• DJIA (no significant relationships)• S&P500
• NASDAQ
Results for Granger causality DJIALag (days)
1-3 5 8-9Open price
Emotion category:
Fearful
word: FearFear, sad
Emotion category:
worry
Close priceMaxMin
Main results
• Emotions angry, sad, tired are Granger cause for volatility of DJIA, S&P500, NASDAQ
• Emotion fearful – Granger cause Open, Close, Min and Max Prices
Model ARMA(p,q)-GARCH(k,m)
The influence of sentiment words (x-axis) with 95% confidence intervals
Indexes’ volatility, with p, q, k and m ranges from 0 to 3, choosing with best MSE criteria
Further plans
• Testing econometric models of mood influence
• Different market situations (application of Markov-chain models)
• Using non-normalized words, emoticons, interjections
• Compare with realistic (5 minutes) volatility
Thank you for attention,