Stock Sentiment from Tweets — TM 2025 (Group 19)

NOVA IMS — MSc in Data Science & Advanced Analytics · Text Mining 2025

Predict market sentiment (bearish=0, bullish=1, neutral=2) from tweets about stocks. We benchmark classic ML baselines (Count/TF-IDF + linear models) and a transformer encoder, selecting DistilBERT as the final model (best Macro-F1).

Task & Data

• Train: tweets with text and label (0/1/2) · Test: tweets with text only.
• Deliverables: tm_tests_19.ipynb, tm_final_19.ipynb, pred_19.csv, and a ≤10-page PDF report.
  (See the official handout for details and deadline.)

Approach

• EDA: label distribution, token length, top words/bigrams after normalization.
• Preprocessing: regex cleaning, lowercasing, stopword removal, lemmatization/stemming.
• Features: TF-IDF (unigrams–trigrams), transformer embeddings (DistilBERT).
• Models: LinearSVC / Logistic Regression baselines; DistilBERT classifier (final).
• Evaluation: Accuracy, Precision, Recall, Macro-F1 on validation.

Results (Macro-F1)

• DistilBERT (Transformer): 0.82 ← final model
• TF-IDF (1–3) + LinearSVC: 0.73
• TF-IDF (1–2) + LinearSVC: 0.73
  (See reports/All_Models_MacroF1_Table.docx for the full table.)

How to reproduce

Open notebooks/tm_final_19.ipynb and run all cells to produce submission/pred_19.csv.

Repository layout

notebooks/ — notebooks (tm_tests_19.ipynb, tm_final_19.ipynb)
reports/ — Text Mining Report.pdf, All_Models_MacroF1_Table.docx, Project Handout TM 2025 v2.pdf
submission/ — pred_19.csv, optional metrics

Authors — Group 19

André Oliveira · Diogo Andrade · Francisco Pontes