tweetstock: Regression Analysis from Public Tweet Sentiment toward NASDAQ stock prices

Given tweets about NASDAQ top 6 stocks(AAPL, GOOG, GOOGL, TSLA, AMZN, MSFT), will be there any relationship between tweet sentiments and the stock price?

Directory structure

|- data
  |- company-sentiment
    : sentiment labeled tweet ids on NASDAQ stocks
  |- raw
    |- company-tweets.csv
    |- company-values.csv
    |- nasdaq-tweets.csv
    |- (sentiment labeled financial tweets)
  |- regression
    : data for regression of public tweet sentiment and market value
  |- sample
    : data for debugging data generator
  |- sentiment-classifier
    : data to finetune BERtweet for sentiment classification of financial tweets
|- calculate.py
  : calculate public sentiment and stock price difference of nasdaq stocks for intervals 
|- finetune_classifier.py
  : finetune model from `fintweet_sentiment_classifier.py`
|- fintweet_sentiment_classifier.py
  : model copied & slightly modified from HuggingFace Transformer RoBERTa
|- nasdaq_tweet_sentiment_tagger.py
  : tag sentiment to nasdaq tweets with fine-tuned classifier
|- preproces_tweets.py
  : drop duplicated and suspicious nasdaq tweets and sentiment labeled financial tweets
|- regression.py
  : output the relation between public sentiment and stock price direction
|- utils.py
  : misc functions

Experiement Pipeline

Step 1. Preprocess tweets to get rid of spam

• invovles preproces_tweets.py
• Drop duplicated or suspicious spam tweets
• Spam filter was based on [Kaggle notebook of aramacus] [1]
• Due to limitation of computing power, sampled 30% of the tweets

  Step 2. Finetune BERtweet with sentiment labeled financial tweets

• invovles fintweet_sentiment_classifier.py, finetune_classifier.py
• To make sentiment classifier for financial tweets, finetune BERtweet (which is RoBERTa pretrained on tweeter data) with sentiment labeled financial tweets

  Step 3. Caculate public sentiment & stock price difference of NASDAQ tweets

• invovles nasdaq_tweet_sentiment_tagger.py, calculate.py
• Tag sentiment to NASDAQ tweets
• Calculate public sentiment for 3 days
• Calculate price difference of:
  • open price of the day right before the duration
  • close price of the day right after the duration

    Step 4. Regression

• involves regression.py
• Output the relation between public sentiment and stock price direction
• Trained for tweets of 2015/01/01-2017/12/31
• Tested for tweets of 2018/01/01-2019/12/31

Results

*All values are accuracy of predicting rise or fall

Discussion

References

[1]: https://www.kaggle.com/aramacus/bot-hunting-or-how-many-tweets-were-made-by-bots
[2]:
@inproceedings{bertweet,
title = {{BERTweet: A pre-trained language model for English Tweets}},
author = {Dat Quoc Nguyen and Thanh Vu and Anh Tuan Nguyen},
booktitle = {Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations},
year = {2020},
pages = {9—14}
}