Skip to content

Latest commit

 

History

History
61 lines (40 loc) · 3.34 KB

MVP.md

File metadata and controls

61 lines (40 loc) · 3.34 KB
Raw

MVP

Sabrina Yang

After collecting $TSLA tweets from 07/12 - 07/19 (total 18019 tweets), I cut down to 10096 tweets after deleting duplicates rows/tweets.

1. Topic Modelings - NMF

Since tweets are considered as relatively short documents, I chose NMF as my method on analyzing the interpretable topic models. Here’s what I found:

  • topic_0 : advertisement - promoting trading alert chatroom
  • topic_1 : FUD (Fear ,Uncertainty, Doubt) # insolvency = bankrupty # crash, panic
  • topic_2 : advertisement - stock app promotion
  • topic_3 : advertisement - stock option trading investment advice
  • topic_4: how much Tesla made from bitcoin investment
  • topic_5: car, FSD Subscription (ha = have: car ownership)
  • topic_6: # fintwit # wallstreetbet # trends # top stocks: aapl, spce(Virgin Galactic), mrna
  • topic_7: advertisement - daily alert
  • topic_8: advertisement - sign up
  • topic_9: stock technical analysis and price charts
  • topic_10: advertisement - stock alert for discord
  • topic_11: meme stock - influencers account trade ideas - promoting certain stocks - amc, gme
  • topic_12: elon musk tweets activity, solarcity, cybertruck, trial, court (Tesla bought Solarcity recently)
  • topic_13: TSLA stock personal opinions and feelings
  • topic_14: TSLA stock personal trading ideas - Short > Long ("Short" shows up more frequent than "Long", means the market is more bearish at that moment)

What do the topics tell you about the overall structure of the data?

  • some of topics are noises ( daily alert, promotions, advertisements) but some are useful

    • elon musk's tweets
    • tech chart patterns
    • retail traders about tesla news

2. Sentiment Analysis Score - Vader

The reason I chose NLTK Vader for this part is because Vader has better sense to pick up social media sentiments than other packages. It comes as 4 scores: negative, neutral, positive and compound. For this case, I will focus on compound score to make a scatterplot based on daily average sentiment score.(why did I take compound score, not other scores to analyze? Since neutral tweets do not reflect a positive or negative mood and serve therefore no purpose to this analysis)

  • a. Daily Sentiment Compound Score

  • b. Daily $TSLA price and Sentiment Compound Score
    • Can't see any significant correlation pattern but the sentiment did fall down when the stock price dropped for the first 2 days(7/12-7/14).
    • Sentiment climbed up and down repeatly from 7/15 to 7/19 even though the stock didn't recover during last week.
    • In the end of this observation timeframe, the sentiment hyper somehow boost the confidence of stock holders and we saw the price recovery on 7/20.

Next step:

  1. look into elon musk's tweets content

  2. do more data cleaning on advertisement tweets to decrease the noises and redo the topic modeling to get clearer topics.

    • update: cut down to 9712 tweets after this cleaning

  3. do Kmeans clustering for topic-matrix df for topic purpose

  4. Check other related tweets than my current keyword "$TSLA" search, for example, check "Tesla" tweets.