Within-call sentiment variance as a market-neutral signal

earnings_call_tone_research/
│
├─ run_backtest.py            # one-shot: build → neutralise → weights → PnL
├─ src/                       # pure-Python research pipeline
│   ├─ load.py                # reads the three parquet inputs
│   ├─ factor_build.py        # z-scores tone dispersion by trade-date
│   ├─ neutralise.py          # FF5+UMD regression (daily)
│   ├─ portfolio.py           # weight construction & PnL
│   └─ report.py              # quick tear-sheet
│
├─ data/                      # **exactly three parquet files**
│   ├─ tone_dispersion.parquet   # call-level dispersion
│   ├─ stock_prices.parquet      # wide adj-close matrix, index=date
│   └─ ff5_daily.parquet         # daily MKT, SMB, HML, RMW, CMA, UMD, RF
│
└─ tests/                     # pytest suite (unit + integration)

No other data sources or credentials are required.

The sentiment analysis component of this research relies on earnings call transcripts. The specific dataset used for building or applying the sentiment models can be found at:

• Kurry's S&P 500 Earnings Transcripts on Hugging Face

• Granularity: Each speaker-turn (a continuous block of speech by one speaker) inside the earnings-call transcript is scored.

• Sentiment score: Each turn receives a scalar in [-1, 1] from a local language-model sentiment head (or any deterministic sentiment function):

  sentiment(turn_i) → s_i
  

• Dispersion metric: The population variance of all sentiment scores within the call (including both executive and analyst turns):

  $$ \text{tone_dispersion} = \frac{1}{N} \sum_{i=1}^{N} (s_i - \bar{s})^2, \qquad \bar{s} = \frac{1}{N} \sum_{i=1}^{N} s_i $$

  Requires at least two non-empty turns; otherwise, the value is NaN.

  Note: This formula requires a Markdown viewer that supports LaTeX math rendering. For reference, in plain code:

  mean = sum(scores) / len(scores)
  variance = sum((s - mean) ** 2 for s in scores) / len(scores)

The resulting table written to data/tone_dispersion.parquet therefore holds one row per call:

During factor construction the script:

1. Shifts date to the next NYSE trading day (trade_date).
2. Z-scores tone_dispersion cross-sectionally each trade_date to remove level shifts.
3. Feeds the standardized series to the FF5+UMD neutralisation step.

# 1. create environment
python -m venv venv
source venv/bin/activate
pip install -r requirements.txt

# 2. drop the three parquet files into data/

# 3. run the pipeline
python run_backtest.py          # writes outputs/weights.parquet & tear-sheet

# 4. run tests
pytest -q                       # sanity-checks factor, alignment and PnL

You can host the backtest results and tear sheet using GitHub Pages:

1. Run python run_backtest.py to generate the outputs/tearsheet.html file.
2. Copy the generated tear sheet to the docs/ directory or reference it from there.
3. Commit and push the docs/ directory to your repository.
4. In your GitHub repository settings, enable Pages with:
   • Source: main branch
   • Folder: /docs
5. Navigate to https://kurry.github.io/earnings_call_tone_research/ to view the documentation site.

This project is licensed under the MIT License. See the LICENSE file for details.