PC Lab 3 — Creating a Factor from Text Data

The setting

Word travels fast. This time it’s a sophisticated hedge fund, and the manager has been watching single tweets move billion-dollar valuations.

“I want to know if Twitter has predictive power. Take this sample of financial tweets: run a proper sentiment analysis on it — and then build me a firm-level measure of media attention. If firms the internet talks about earn different returns… that may be a great factor idea.”

Your goals: clean real text data (messier than any price series you’ve met), do sentiment analysis with modern NLP, and take a first shot at factor construction — the bridge from last week’s lecture to the data-driven finance of this one.

Background

This week’s lecture, Empirical Asset Pricing, covers factors, how they’re built, and the factor zoo. For the text part you’ll meet new tools: nltk, ekphrasis, word clouds, and pretrained transformers (FinBERT, Twitter-RoBERTa) or LLM-based classification.

The data

Data_PCLab3_Twitter_Stock_Sentiment.csv: real financial tweets, two columns — Text (the tweet) and Sentiment (a 0/1 label of unknown quality; part of the fun is deciding whether to trust it). For the optional factor part, the daily Fama–French 5 factors file is provided too.

Your assignment

Part 1 — Know your sample (Task 1)

• Import the tweets and describe the sample: how many tweets? Words per tweet and their distribution? Average sentiment?

Part 2 — Clean and visualize (Task 2)

• The usual text-cleaning pipeline: remove URLs, mentions, hashtags, emojis; tokenize, lemmatize, drop stopwords. For tweet-specific quirks (elongated words, emoticons, hashtags) try ekphrasis.
• Plot a word cloud for positive and negative tweets separately.
• How many unique words are left?

Part 3 — Sentiment analysis (Task 3)

• The dataset ships with a sentiment label — should you trust it?
• Produce your own sentiment with either:
  • a pretrained transformer (e.g. Twitter-RoBERTa, FinBERT), or
  • an LLM (zero-shot / few-shot classification).
• Compare your classification to the original labels: how do these methods perform on financial tweets?

Part 4 — Measuring media attention (Task 4)

• Using the ticker list from last week’s scraping exercise, count tweets per stock and rank stocks by media attention — total, positive vs. negative, disagreement… be creative.
• Build 10 portfolios sorted on your preferred attention measure: is there a correlation between media attention and returns? Could Twitter attention be a factor?
• Optional: regress each stock’s excess returns on the Fama–French 5 factors, keep the residuals (the abnormal returns traditional factors can’t explain), and redo the plot with residuals. Does attention explain what FF5 cannot?

Deliverables

• One Jupyter notebook per group, by email to the instructor (clement.mazetsonilhac@unibocconi.it) and the TA (andrea.andolfatto@phd.unibocconi.it).
• Email title — and notebook name: PCLab#3 - Group X - Name1 Name2 Name3. Tell us how long it took.
• Deadline: Friday, midnight (end of Week 4).

Grading

1. Submitting before the deadline;
2. Code quality — comments, readability, use of functions;
3. Notebook structure — well organized, explain what and why;
4. Completing the tasks and innovating — produce less, but more useful output.