Efficient Market Hypothesis

• Stock prices represent all available information
• 3 Variants, “weak”, “semi-strong”, and “strong”
• Strong form contends that prices reflect even inside information (information that is illegal to use in trading)
• Under semi-strong, beating the market impossible for law abiding people

Why Bother?

• There is much controversy over the truth of EMH
• At its core, EMH is about information
• Gaining an informational advantage could allow you to “beat” the market
• Twitter is a vast source of largely untapped data about people’s opinions of companies
• Mining the Twitterverse for these opinions could be of comparable value to “insider info”

Measuring Opinions

• Sentiment analysis seeks to quantify opinions based on text data
• Many forms, e.g. statistical, knowledge based, machine learning
• Knowledge based starts with a set of words, and each word is assigned a hand coded value
• With the vocabulary I used, AFINN, each word has a score between -5 and 5
• The sentiment associated with an opinion is the sum of the AFINN based scores of its words

Harvesting Tweets

• Free Twitter API
• Allows user to search for tweets related to companies
• Each tweet about a company was assumed to be an opinion regarding that company and its sentiment score was measured
• Tweets were stored using another tool having a free version, Mongo Atlas
• Data was harvested daily

Companies Chosen

Top Words

Sentiment Evaluation

Testing with hand coded Set:

Number positive: 0.4652088
Accuracy: 0.6475034

Sentiment Evaluation

Stock Price Vs Sentiment

• Sentiment varied wildly between companies
• Stock price changes were on different scale than sentiment
• Both variables standardized

Does it Work?

• Stock price and previous day’s sentiment are uncorrelated

What now?

• Continue to improve ability to assess the sentiment of tweets
• More stock prices plus sentiment data points will provide more options for modeling techniques
• Hand code some data
• Pivoting always possible if the sentiment model continues to improve
• Many companies would pay to know consumer’s sentiments regarding their brands and products