Transformer-based NLP pipeline analyzing 77k earnings-call transcripts to link corporate tone with short-term stock returns.

This project builds an end-to-end NLP pipeline that analyzes 77k corporate earnings-call transcripts (2010–2025) to quantify tone and emotion using transformer-based models (FinBERT + Emotion DistilRoBERTa).
Developed as an independent research project at TU Dortmund University, the system aggregates token-level outputs into interpretable call-level sentiment features and links them with short-term post-call stock returns through regression analysis. Results show that tone variables are individually highly significant yet they only explain a small share of return variance, which is consistent with market efficiency. More details below.

┌────────────────────────┐        ┌────────────────────────────┐        ┌───────────────────────────┐
│ 1                      │        │ 2                          │        │ 3                         │
│ Data Preprocessing     │ ─────► │ Transformer Inference      │ ─────► │ Feature Engineering       │
│                        │        │                            │        │                           │
│ • Clean + normalize    │        │ • Sentim. & Emotion models │        │ • Aggregate token probs   │
│ • Sentence splitting   │        │ • Batched inference (GPU)  │        │ • Weighted means, entropy │
│ • Chunk to 512 tokens  │        │ • Cache intermediate reps  │        │ • Dispersion, extremes    │
└────────────────────────┘        └────────────────────────────┘        └───────────────────────────┘
                                                                                           │
                                                                                           ▼
┌─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ┐       ┌────────────────────────────┐        ┌────────────────────────┐
│ 6                          │       │ 5                          │        │ 4                      │
│ Visualization & Analysis   │ ◄──── │ Regression Modeling        │ ◄───── │ Market Data Alignment  │
│                            │       │                            │        │                        │
│ • Correlations, deciles    │       │ • OLS, statsmodels         │        │ • Merge tone + prices  │
│ • Coeff. trends, heatmaps  │       │ • Evaluate R², RMSE        │        │ • Compute 1d/3d/5d     │
│ • Interpret market signal  │       │ • Combined Sentim.+Emotion │        │   forward returns      │
└ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─┘       └────────────────────────────┘        └────────────────────────┘

Frameworks

PyTorch · Hugging Face🤗 Transformers · spaCy · pandas · pyarrow · scikit-learn · statsmodels · matplotlib · seaborn

Models

• FinBERT (ProsusAI/finbert) – financial sentiment classification (positive / negative / neutral).
• Emotion DistilRoBERTa (j-hartmann/emotion-english-distilroberta-base) – 7-way emotion classification (anger, disgust, fear, joy, neutral, sadness, surprise).

Regression analyses were conducted over 62k matched earnings-call transcripts and stock-return observations to evaluate how linguistic tone relates to short-horizon market reactions.
Three models were tested:

• FinBERT Sentiment,
• DistilRoBERTa Emotion tone,
• Combined model integrating both feature sets.

Directionally consistent effects
Positive tone and emotional valence are associated with higher post-call returns, while negative sentiment and emotions (e.g., sadness, fear) correlate with lower returns.

Short-term concentration Predictive strength peaks at the 1-day horizon and decays over 3–5 days, indicating that markets rapidly incorporate tone information following earnings calls.

Complementary signals
The combined model outperforms sentiment or emotion features alone, suggesting that emotional nuance adds incremental explanatory power to finance-specific sentiment.

Statistical strength, modest economic impact
Although coefficients are highly significant (t-stats >15 across tone variables), overall explanatory power remains small (R² ≤ 0.013). This is consistent with prior literature on financial-text analytics.

Interpretation
Results confirm that corporate communication tone carries short-lived yet measurable predictive information, reinforcing the role of NLP-derived sentiment and emotion features as auxiliary signals in quantitative trading and event-driven models.

This analysis isolates the linguistic component of market reactions, focusing on tone–return causality rather than price levels or the basics.

Future work could extend this by:

• Incorporating firm and sector fixed effects to control for baseline performance.
• Modeling interaction terms between tone and earnings surprises.
• Testing temporal dynamics (rolling regressions, expanding windows) for stability.
• Exploring multi-modal features (audio tone, speaker sentiment, Q&A sections).

# Clone the repository
git clone https://github.com/aryahassibi/Earnings-Call-Sentiment-and-Stock-Returns.git
cd Earnings-Call-Sentiment-and-Stock-Returns

# Create and activate a virtual environment
python3 -m venv .venv
source .venv/bin/activate  # macOS / Linux
# .venv\Scripts\Activate.ps1  # Windows

# Install dependencies
pip install -r requirements.txt
# update reqs with pip freeze > requirements.txt

Optionally, install the spaCy language model

python -m spacy download en_core_web_sm

and register a Jupyter kernel for the virtual environment to appear in the Jupyter UI

# Register Jupyter kernel
python -m ipykernel install --user --name earnings-call-sentiment --display-name "Earnings Call Sentiment"

Run python utils/analyze_parquet.py <parquet_file> to generate a JSON profile and Markdown summary for any dataset. Use --help for options (sampling, alternate text columns, pandas vs. pyarrow engine).

process_var_description_docx.py extracts variable descriptions from the docx file and saves them as a Parquet file.

• 00_data_overview.ipynb: dataset exploration, health checks, and transcript statistics.
• 01_sentiment_scoring.ipynb: production-ready sentiment feature pipeline with caching.
• 01v2_sentiment_scoring.ipynb: Runs FinBERT ( or yiyanghkust/finbert-tone) on transcripts; outputs call-level sentiment features (including prepared/Q&A and speaker-role deltas).
• 02_returns_regression.ipynb: merge sentiment features with market data and fit baseline regressions.
• 03_emotion_analysis.ipynb: Applies the DistilRoBERTa emotion classifier to transcripts, explores emotion distributions, and merges with returns.
• 04_sentiment_returns_regression.ipynb: Regression suite testing FinBERT and DistilRoBERTa features against post-call returns (single-variable, multivariate, etc.).

• data/transcripts.parquet: Raw earnings-call transcripts with symbol, year, quarter, and the full transcript text.
• data/ec_mapping.parquet: Maps each transcript (symbol, year, quarter) to the earnings-call date rdq used to align with market data.
• data/algoseek_nyse_nasdaq.parquet: Daily OHLCV data (including CloseAdjusted) for NYSE/NASDAQ tickers, used to compute post-call returns.
• data/additional_ec_data.parquet: Quarterly fundamentals keyed by ticker (tic) and event date (rdq_x); useful for control variables.
• data/var_descriptions.parquet: Metadata describing the variables in the additional fundamentals dataset.
• output/sentiment_features*.parquet / output/emotion_sentiment_features.parquet: Feature tables produced by the scoring notebooks (FinBERT and DistilRoBERTa emotion models).