Earnings Sentiment & Price Reaction Analyser

An NLP pipeline that scores earnings call transcripts using domain-specific financial sentiment lexicons and tests whether management tone predicts short-term abnormal equity returns — replicating the methodology of academic finance NLP research.

---

Research question

Does positive (negative) sentiment in earnings call transcripts predict positive (negative) abnormal stock returns in the 1, 3, and 5 trading days following the call?

---

Pipeline architecture

SEC EDGAR 8-K filings + synthetic fallback
              │
              ▼
┌─────────────────────────┐
│  1. Transcript          │  SEC EDGAR API → text extraction
│     Collection          │  → synthetic fallback for unavailable filings
│     fetcher.py          │
└──────────┬──────────────┘
           │  raw transcript text
           ▼
┌─────────────────────────┐
│  2. Sentiment           │  Loughran-McDonald financial lexicon
│     Analysis            │  + VADER rule-based scorer
│     analyser.py         │  → composite score + label
└──────────┬──────────────┘
           │  sentiment scores per earnings event
           ▼
┌─────────────────────────┐
│  3. Price Reaction      │  Event study: abnormal returns
│     Analysis            │  → correlation test + t-test
│     price_reaction.py   │  → interactive dashboard
└─────────────────────────┘

---

Methods

Sentiment scoring

Two complementary approaches are combined into a composite score:

Loughran-McDonald (LM) Financial Lexicon Domain-specific wordlist calibrated for financial text. Words like "uncertainty" and "risk" are negative in this context — unlike general-purpose sentiment models which treat them as neutral. Standard in academic finance NLP. Weight: 60% of composite score.

VADER (rule-based) Handles negation ("not profitable"), intensifiers ("significantly exceeded"), and capitalisation. Fast and interpretable baseline. Weight: 40% of composite score.

Composite score: 0.6 × LM_net + 0.4 × VADER_compound ∈ [-1, 1]

Event study

• Abnormal return: AR(t) = R_stock(t) - R_market(t) (market-adjusted model)
• Cumulative abnormal return: CAR[1,N] = Σ AR from t+1 to t+N
• Event windows: 1-day, 3-day, 5-day post-earnings
• Statistical tests: Pearson correlation + independent samples t-test (positive vs negative sentiment groups)

---

Modules

Module 1 — Transcript Collection (1_collection/fetcher.py)

• Queries SEC EDGAR submissions API for recent 8-K filings
• Extracts text from filing documents
• Falls back to deterministic synthetic transcripts where SEC data unavailable
• Outputs: output/transcripts/transcripts.csv

Module 2 — Sentiment Analysis (2_sentiment/analyser.py)

• Scores each transcript with LM lexicon and VADER
• Computes composite score and categorical label (positive/neutral/negative)
• Outputs: output/sentiment/sentiment_scores.csv, ticker_sentiment_summary.csv

Module 3 — Price Reaction Analysis (3_analysis/price_reaction.py)

• Joins sentiment scores to forward abnormal returns
• Runs correlation analysis and t-tests across event windows
• Generates interactive Plotly dashboard
• Outputs: output/analysis/event_study_results.csv, statistical_summary.csv, sentiment_analysis_dashboard.html

---

How to run

1. Install dependencies

pip install -r requirements.txt

2. Run the full pipeline

python run_pipeline.py

3. View results

open output/analysis/sentiment_analysis_dashboard.html

Or run modules individually:

python 1_collection/fetcher.py
python 2_sentiment/analyser.py
python 3_analysis/price_reaction.py

---

Outputs

File	Description
output/transcripts/transcripts.csv	Raw transcript text per ticker per quarter
output/sentiment/sentiment_scores.csv	Per-transcript LM, VADER, and composite scores
output/sentiment/ticker_sentiment_summary.csv	Aggregated sentiment per ticker
output/analysis/event_study_results.csv	Sentiment joined to forward CAR
output/analysis/statistical_summary.csv	Correlation and t-test results
output/analysis/sentiment_analysis_dashboard.html	Interactive dashboard

---

Dependencies

yfinance>=0.2.28
pandas>=2.0.0
numpy>=1.24.0
scipy>=1.10.0
plotly>=5.0.0
requests>=2.28.0

---

Limitations and extensions

Known limitations:

• Synthetic transcripts are used where SEC filings are unavailable — results using synthetic data should be interpreted as methodology demonstration only, not empirical findings
• Simple market-adjusted abnormal returns (no beta estimation)
• No correction for multiple testing across windows

Natural extensions:

• Replace synthetic transcripts with a paid transcript provider (Refinitiv, Bloomberg, Motley Fool)
• Add FinBERT transformer-based scoring for comparison
• Incorporate conference call Q&A section separately from prepared remarks
• Extend to analyst report sentiment

---

Author

Atrija Haldar LinkedIn MSc Engineering, Technology and Business Management — University of Leeds