Skip to content

Interactive app

The Streamlit app lets you change the parameters and watch the estimates move, rather than reading fixed numbers off a page.

Open the live app :material-open-in-new:

Or run it yourself:

pip install -e ".[app]"
streamlit run app/streamlit_app.py

The app fetches the dataset itself on first run if there isn't one on disk, so there's no separate setup step. (Running python -m quantile_compass.fetch_data beforehand just gets it out of the way sooner.)

What you can change

Control Effect
Confidence level The VaR quantile, 90% to 99.9%.
EWMA decay (λ) How fast the volatility estimate forgets. Lower λ reacts faster to shocks and is noisier; higher λ is smoother and slower.
Horizon Scales the 1-day figure by √h.
Portfolio weights and betas Re-runs the whole decomposition — the split between equity and currency risk shifts with them.

What it shows

  • Volatility — the two EWMA series with crisis windows shaded.
  • Correlation — the equity/forex correlation through time.
  • Return distribution — the histogram with the parametric and historical VaR thresholds drawn on it, so you can see where each method puts the cut-off.
  • Risk factors — the supplementary series (S&P 500, VIX, gold) for context.

Things worth trying: drop λ to 0.80 and watch the volatility series get much choppier; push the confidence level to 99.9% and see the parametric and historical estimates diverge further as the normal assumption strains in the deep tail.

Deploying it

The app is ready for Streamlit Community Cloud as-is:

  1. Sign in with GitHub and pick New app → this repository.
  2. Set the main file path to app/streamlit_app.py, branch main.
  3. Deploy.

Two things make that work without further setup. The root requirements.txt holds only the app's runtime dependencies, so the deploy doesn't drag in the test, notebook and docs toolchains. And the app puts src/ on sys.path itself, so it imports quantile_compass correctly even though nothing pip-installs the package there.

First load is slower

A fresh deploy has no dataset, so the first visit spends a moment fetching ~20 years of daily history before anything renders. It is cached from then on, until the container restarts.