Twitter Sentiment & Investing - modeling stock price movements with twitter sentiment.

5,577 views

Published on

In this presentation, I provide an overview of my research into using twitter sentiment and message volume as inputs into modeling stock price movements. A quick and dirty linear regression model using Twitter Sentiment, the Number of Tweets per day, the VIX Closing price and the VIX Price change delivers a simple model for the S&P 500 SPY ETF that has an accuracy of 57% over 6 months (tested on out-of sample data). This model was built using data from July 11 2011 to August 11 2011.

Published in: Economy & Finance, Business
2 Comments
3 Likes
Statistics
Notes
  • However, does it make any sense? Can every trader make tons of money in binary trading? Who is actually paying all the money or the profit to traders? See ANDY LANK METHOD . I am using this product since 3 weeks and i am getting good profit.
       Reply 
    Are you sure you want to  Yes  No
    Your message goes here
  • the worst thing was i blew up my first account. I guaranty you that GINO SHEARER SYSTEM is the best binary trading software. I am over confident because I have also benefited now from Andy Lank Method. ?
       Reply 
    Are you sure you want to  Yes  No
    Your message goes here
No Downloads
Views
Total views
5,577
On SlideShare
0
From Embeds
0
Number of Embeds
672
Actions
Shares
0
Downloads
139
Comments
2
Likes
3
Embeds 0
No embeds

No notes for slide

Twitter Sentiment & Investing - modeling stock price movements with twitter sentiment.

  1. 1. WILL TWITTER MAKE YOU A BETTERINVESTOR?A LOOK AT SENTIMENT, USER REPUTATION AND THEIR EFFECT ON THE STOCKMARKETEric D. Brown
  2. 2. Background• Sentiment has long been an underlying factor in the investing world • Consumer Confidence Index • Investors Intelligence Sentiment Index • “Market Sentiment”• Rather than waiting days, months or weeks, can the „sentiment of now‟ be used to improve trading performance and investing decisions?• Can Twitter be used to determine the „sentiment of now‟?
  3. 3. BackgroundThe thoughts driving this research are:• Can analysis of publicly available Twitter Messages provide insight for decision making for investing?• Do Twitter messages (and their subsequent sentiment) have any effect on movement in the stock market?• Can Twitter messages be mined and analyzed to predict movements in the stock market?• Does a Twitter user‟s reputation have an effect on how people perceive and use their shared investing ideas?
  4. 4. Research Method Stock Twitter Data Market Collection Data Price & Sentiment Social Volume Analysis Analysis Analysis Positive Correlation of sentiment Reputation of and message volume with Twitter user price/volumeUnderstanding of predictive capabilities of Twitter Sentiment and the affect of user reputation investing decision support
  5. 5. Research Method• Data Collection • Using Twitter API to collect tweets (tweet, sender, date, time) • Tweets referencing companies and sectors are collected and stored in a MySQL database for future study • Using the nomenclature made popular by StockTwits (www.stocktwits.com). Example: The stock symbol for Apple is AAPL. Users following the StockTwits nomenclature add a “$” to the symbol – “$AAPL”. • Stocktwits.com describes their purpose as a place to: • …share ideas, market insights and trades on stocks, futures and the market in general *. • Using Yahoo Finance data feed to gather Stock Market data (price and volume) • Provides historical data
  6. 6. Research Method• Sentiment Analysis• Using a Naïve Bayesian text classification algorithm to determine sentiment of collected Tweets • Naïve Bayesian is being used for simplicity but also because many researchers have pointed out very minor differences between it and other sentiment analysis methods • A subset of the data collected has been manually assigned „sentiment‟ to build the necessary training dataset• Using the Python Natural Language Toolkit, the Bayesian classification is performed• For each tweet, the overall score is calculated and assigned. • Ideally, tweets will fall into +1 (Bullish), 0 (Neutral), -1 (Bearish) buckets.
  7. 7. Sentiment Classification Process Bayes Training Dataset Classifier Trained Twitter Data Classifier Classified Twitter Messages
  8. 8. Current Dataset• Twitter Dataset: • May 1 2011 to Dec 31 2011 • 473,901 Tweets • No deduplication performed• Training Dataset: • 5000 messages randomly selected from collected Tweets • Messages have been manually coded as Bullish, Bearish, Neutral or Spam • 544 Bullish (10.88%) • 638 Bearish (12.76%) • 3454 Neutral (69.08%) • 364 Spam (7.28%)
  9. 9. Bullish Examples• Markets seems like consolidating before another rally higher $SPX $SPY $QQQ• $KFT - Kraft Foods Stock Analysis - CCI is bullish and rising• If this daily candle ends like this do not go short! 3 white soldiers (bullish) $SPX $SPY http://t.co/oWWUqnu• buy the dips WORKING :-) $ES_F $SPY
  10. 10. Bearish Examples• $K - Kellogg Stock Analysis - bearish stochastic crossdown• $SPY - Might be trying to roll a little.• warning sign as $IWM didn‟t make a higher high, unlike $QQQ and $SPY• RT @grassosteve: $SPX Jobs although important only 1 aspect of the weakness in the mkts, i would sell pops still levels up 1216 1228 123 ...
  11. 11. Neutral Examples• Stocks and Inflation: What the Market Is Really Telling Us http://seekingalpha.com/a/5tap $TLT $TLH $SPY• NEW POST: UPDATED- MEAN REVERSION TRADE ON THE RUSSELL 2000 http://bit.ly/l2Q1Rf $IWM $TNA $SPY $QQQ• Sold my $SPY Jun 03 2011 133.0 Puts for 78c made 7c• Durable Goods as a Leading Market Indicator http://seekingalpha.com/a/5u8j $DIA $SPY $QQQ
  12. 12. Spam Examples• Hop up out the bed turn my $wag on. ;*• MILLIONAIRE SECRETS CLUB - MAKE $1 MILLION A YEAR http://goo.gl/2Yv8s $OXY $PBR $PDCO $pennystocks $POT $PRU $QCOM $QSII axsc• HOME TYPERS NEEDED - MAKE $1000s WEEKLY - PAYS DAILY http://goo.gl/hoNUn $SNP $SOHU $SPLS $SPRD $SPX $SPY $STO $stocks sg2• UNLIMITED FREE TV SHOWS on YOUR PC - 12,000 FREE CHANNELS http://goo.gl/v55Nw $CBG $CBS $CF $CLR $CMCSA $COF $COP $CROX $CTIC qika
  13. 13. Sentiment Analysis using full training Set• From May 1 to Dec 31 2011: All Tweets 473,901 Bullish 103,770 21.90% Bearish 84,454 17.82% Neutral 224,300 47.33% Spam 61,371 12.95% None 6 0.001%
  14. 14. A look at the Market & Sentiment• I ran a short analysis on the S&P 500 ETF (SPY) between July 11 and August 11 2011 • This date range chosen mainly due to a very volatile movement down • 32 days of data • 26,307 tweets • On Jul 11 SPY is 131.40 • On Aug 11 SPY is 117.33
  15. 15. S&P 500 ETF (SPY) - 7/11 to 8/11
  16. 16. S&P 500 ETF (SPY) - 7/11 to 8/11
  17. 17. S&P 500 ETF (SPY) - 7/11 to 8/11
  18. 18. S&P 500 ETF (SPY) - 7/11 to 8/11
  19. 19. But…is the Classifier accurate?• We can determine sentiment of a tweet…but is it really accurate?• The Python Natural Language Toolkit provides a method to determine accuracy of training dataset • Build training dataset as normal • Use training dataset as the “input data” • Run all messages through classifier and determine accuracy
  20. 20. Sentiment Classification Process Bayes Training Dataset Classifier Trained Twitter Data Classifier Classified Twitter Messages
  21. 21. Classification Accuracy Bayes Training Dataset Classifier Trained Training Dataset Classifier Accuracy
  22. 22. Accuracy of Training Dataset• If you recall, our training dataset is: • 5000 messages randomly selected from collected Tweets • Messages have been manually coded as Bullish, Bearish, Neutral or Spam • 544 Bullish (10.88%) • 638 Bearish (12.76%) • 3454 Neutral (69.08%) • 364 Spam (7.28%) • Running the accuracy method of the Python NLTK delivers a 54.18% accuracy.
  23. 23. How can we improve accuracy?• If we think about the theory behind the Bayesian classifier, it may shine some light on the inaccuracies.• The Bayesian Classifier is a probability based theory and is only as good as the data used to train.• Research suggests that having non-symmetric training data sets / features, can throw the Bayesian filter off.• The training dataset used is non-symmetric.• What if we create a symmetric dataset with the same number of Bullish, Bearish, Neutral and Spam data?
  24. 24. Equivalent Sized Training Dataset• Rebuild Training Dataset• Randomly select 500 tweets from each training dataset• Re-run the accuracy method again. • Accuracy = 91.94%• An improvement from 54.18% to 91.94%• What will this improved accuracy do for the overall dataset?
  25. 25. Sentiment Analysis using Symmetric training Set• From May 1 to Dec 31 2011: All Tweets 473,901 Bullish 110,141 23.24% Bearish 103,233 21.78% Neutral 212,509 44.84% Spam 48,012 10.13% None 6 0.001%
  26. 26. S&P 500 ETF (SPY) - 7/11 to 8/11
  27. 27. S&P 500 ETF (SPY) - 7/11 to 8/11
  28. 28. S&P 500 ETF (SPY) - 7/11 to 8/11
  29. 29. S&P 500 ETF (SPY) - 7/11 to 8/11
  30. 30. Statiscally Speaking• There seems to be some correlation between twitter & Stock data• Correlation of TransformedBBN and Closing Price • Correlation=0.495 • P-Value = 0.004• Correlation of Num Tweets and Daily Volume • Correlaton = 0.648 • P-Value = 0.000
  31. 31. Correlations?• Basic Analysis shows some correlation between price/sentiment and volume/tweet volume.• Using Time Series analysis, a cross-correlation analysis can be completed to determine how these variables are related at different „lag‟ periods.• Using Cross-correlation analysis we can get the Cross- Correlation Coefficient (CCF) which gives us an idea of how well two variables are correlated at lag time r.• If a correlation is found a negative lag time r, that variable is a candidate for use in predicting the output variable.
  32. 32. Correlations• Closing Price Cross Correlation Coefficient Lag Time r Variable -6 -5 -4 -3 -2 -1 0 Sentiment 0.299 0.399 0.599 0.620 0.663 0.599 0.495 Num Tweets -0.497 -0.461 -0.406 -0.533 -0.577 -0.627 -0.714• Volume Cross Correlation Coefficient Lag Time r Variable -6 -5 -4 -3 -2 -1 0 Sentiment -0.307 -0.394 -0.527 -0.574 -0.617 -0.605 -0.557 Num Tweets 0.373 0.443 0.485 0.552 0.613 0.642 0.648
  33. 33. What now?• Correlations exist between price/volume and sentiment/message volume.• The main reason I‟m looking at sentiment is to determine if it can somehow be used to predict price movement.• So…let‟s build a model using Linear Regression (note…linear regression isn‟t a likely fit but a good place to start).• I want the model to predict Closing Price…so let‟s start with a simple model using Sentiment only
  34. 34. Building Models• Using Sentiment to predict Price: • The regression equation is • Closing Price = 133 + 47.3 TransformedBBN • The p-value is less than 0.05, so we should be good but R-Squared is 24.5…which tells me this isn‟t a very good model.
  35. 35. Building Models (cont)• What other variables can be used? • Volume? • Volume change? • Number of Tweets? • Volatility measurements? • There are a lot of combinations…but let‟s keep it simple. Let‟s select the Number of Tweets and re-run the analysis.
  36. 36. Building Models (cont)• Using Sentiment + Number of Tweets:• The regression equation is • Closing Price = 139 + 41.4 TransformedBBN - 0.00856 Num Tweets • P-values look good still and R-Squared is up to 69.6%.
  37. 37. Building Models (cont)• What else can we add?• How about a measure of volatility?• One often quoted measure is the VIX. This is a measurement of implied volatility of S&P 500 index options.• Often referred to as the „fear index‟.• On 7/11, the VIX was 18.39• On 8/11 the VIX was 39.00
  38. 38. Building Models (cont)• Using Sentiment, Number of Tweets, the VIX Closing Price + the change in price for the VIX• The regression equation is • Closing Price = 148 - 0.00197 Num Tweets + 12.1 TransformedBBN - 0.713 VIX Closing Price + 0.128 VIX Price Change • P-values are good & R-Squared hits 97.3%...which tells us that this might be a good model.
  39. 39. Checking the model• Our equation is: • Closing Price = 148 + 12.1 TransformedBBN - 0.00197 Num Tweets - 0.713 VIX Closing Price + 0.128 VIX Price Change• So…for August 11, our variables are: • TransformedBBN = -0.075 • Num Tweets = 1357 • VIX Closing Price = 39 • Vix Price Change = -3.99• Our predicted closing price for August 12 is then: 116.65 • The prediction is for a move down from the August 11 price of 117.33• The actual closing price from August 12 was 118.20 • Our directional prediction was incorrect and our price prediction was incorrect.
  40. 40. Checking the Model• A few more predictions (50% accuracy) :• For August 15 – prediction is up. Price is 118.44 • Actual is 120.62. Price moved up.• For August 16 – prediction is up. Price is 121.69 • Actual is 119.59. Price moved down.• For August 17 – prediction is up. Price is 121.07 • Actual is 119.67. Price moved up.• For August 18 – prediction is up. Price is 121.73 • Actual is 114.51. Price moved down
  41. 41. Checking the model• The model works for prediction direction on a few dates immediately…but about the rest of the time?• Looking at the rest of the year, there are 99 trading days. • For 53 days, the prediction was correct. • For 46 days, the prediction was incorrect.• The model gives a 53.54% accuracy. • Not great….but better than 50%. • With proper risk management of investments, a 3.54% “edge” on the market might be perfectly acceptable.• FYI - This model applied to a full 6 month data set gives an accuracy of 57.06%
  42. 42. Checking the model• As a test, I removed the Tweet Sentiment and the Number of Tweets from consideration and only re-ran an analysis to create a model.• The model uses the VIX and VIX Price change only and gives R-Squared of 95.5%.• The regression equation is • Closing Price = 149 - 0.847 VIX Closing Price + 0.120 VIX Price Change• This model gives an accuracy of 48.48%....so there seems to some value in sentiment/tweet volume.
  43. 43. Next Steps• There seems to be some correlation between twitter & Stock data• Begin building more complex predictive models using Time Series modeling and prediction methods (ARIMA, etc).• Continue analysis of sentiment and price movements.• Begin Social Network Analysis of twitter users for reputation, etc
  44. 44. Thank you Eric D. Browneric@ericbrown.comhttp://ericbrown.com

×