📈 Sentiment-Driven Returns: PRIO3 (PetroRio) 2015 – 2024

Technologies │ Python • Pandas • SpaCy • Transformers (FinBERT-PT-BR) • yfinance • Matplotlib • Seaborn • Statsmodels • SciPy

---

1 | Introduction

Brazil’s independent oil producer PetroRio S.A. (ticker PRIO3) moved from penny-stock to Ibovespa heavyweight in less than a decade. During that climb its newsflow exploded—earnings beats, field acquisitions, OPEC shocks, regulatory twists. Can the tone of those headlines help explain (or even predict) the stock’s price action?

• Scrape Portuguese headlines (2015-2024)
• Score tone with FinBERT-PT-BR → compound ∈ {-0.5, 0, +0.5}
• Build daily features: compound_mean, compound_mag, art_count
• Join to B3 prices and compute forward returns out to +90 days
• Analyse with event‐study curves, horizon bar-plots, violin distributions, correlation matrices, Granger causality and mean-difference tests.

---

2 | Main take-aways

Finding	Evidence
Intraday impact negligible	Median pct_d0 ≈ 0 % for all sentiment classes.
Medium-term drift appears	Neutral & Positive days beat Negative by 15–30 pp at D + 15 / 30.
Statistical power weak so far	Welch-t / Mann-Whitney on ≤ 7-day horizons not significant (p > 0.05).
compound_mean ≈ compound_mag	Pearson ρ ≈ 0.94.
Sentiment–return link non-linear & lagged	Pearson/Spearman ≈ 0 in short run; effect shows only after 5 + days.

Bottom line: headline tone doesn’t move PRIO3 intraday, yet there is early evidence of bullish drift 5–30 days after non-negative headlines—worth deeper back-testing with more data.

---

3 | Steps to reproduce

# Clone repo
git clone https://github.com/jp-alves/prio3-sentiment.git
cd prio3-sentiment

# 1) create env
conda env create -f environment.yml
conda activate prio3

# 2) download models
python -m spacy download pt_core_news_sm
python -m nltk.downloader vader_lexicon

# 3) scrape & clean
python src/scrape/scrap_news.py
python src/scrape/scrap_prices.py
python src/clean/clean_news.py
python src/clean/clean_prices.py

# 4) NLP + merge
python -m src.nlp.apply_sentiment
python analysis/merge.py

# 5) open notebook
jupyter lab                 # notebook/main_analysis.ipynb