Bandit OPE Tutorial (Script)

This tutorial mirrors the examples/quickstart/bandit_ope.py script. It is designed for researchers who want a minimal, reproducible bandit OPE run that still exposes diagnostics and uncertainty.

Run the script

python -m examples.quickstart.bandit_ope

What it does

1. Generates a synthetic contextual bandit dataset (ground truth known).
2. Defines a PolicyValueEstimand with overlap and ignorability assumptions.
3. Runs IS and WIS.
4. Prints a report table with estimates and diagnostics.

What to look for

• Overlap diagnostics: are importance ratios heavy-tailed?
• ESS: small values suggest unstable estimates.
• IS vs WIS gap: large gaps hint at variance issues.

Next steps

• See the full notebook workflow: notebooks/02_bandit_ope_walkthrough.ipynb
• Use evaluate_ope(...) for multi-estimator reports.