FedLens

Interactive Streamlit dashboard tracking Fed sentiment through embeddings and cosine similarity analysis.

Thesis: Language shifts between FOMC meetings lead actual rate decisions by 1–2 meetings. This dashboard tracks that drift in real time.

What It Demonstrates

• Hawk-dove sentiment scoring via geometric projection onto learned axis
• Language drift detection between consecutive FOMC meetings
• Change point detection (PELT) for regime identification
• Lead/lag correlation analysis (EWM z-score at lag +1, r = 0.47)
• 3D UMAP trajectory visualization

Stack

• Python 3.11
• Streamlit (multi-page navigation)
• Plotly (interactive charts)
• sentence-transformers (all-mpnet-base-v2)
• ruptures (PELT change point detection)

Setup

make install

Run Locally

make run

Or directly:

streamlit run app.py

App will open at http://localhost:8501

Project Structure

app.py                    # Entry point + st.navigation
pages/
  overview.py             # Key findings + alert banner
  sentiment.py            # Hawk-dove score timeline
  drift.py                # Language drift velocity + change points
  trajectory.py           # 3D UMAP trajectory
  methodology.py          # How it works + links
src/
  data/loader.py          # Load + cache FOMC data
  signals/sentiment.py    # Hawk-dove projection (pure functions)
  signals/drift.py        # Cosine drift computation
  signals/changepoint.py  # PELT change point detection
  viz/charts.py           # All Plotly chart builders
data/
  fomc_statements.pkl     # Embedded + scored statements
  hawk_dove_axis.pkl      # Pre-trained hawk-dove axis
  fed_funds_rate.pkl      # Fed funds rate from FRED

Data Files

All data files are pre-computed and should be checked into the repo before deployment. See data/README.md for schemas.

Deployment

Deploy to Streamlit Community Cloud:

1. Connect GitHub repo
2. Select app.py as entry point
3. Deploy (cold start ~30s with sentence-transformers)

Related

• Medium article: (link when published)
• Research notebook: model.ipynb
• Trillion Dollar Words dataset (ACL 2023)