[bradfordlynch]

Summary

• New notebooks/04_options_market_data.ipynb: reads the option-implied-stats item, then uses the implied move as an event-specific scale for a submission's predictions. The score is invariant to a uniform affine remap, so only a per-event factor can move it.
• The naive rescale (pivot 0.5, linear in the move) lowers R-squared for both reference baselines. Pivoting at the submission's own mean with a square-root factor raises it for Gemini (0.103 to 0.108 on covered 2026Q3 events) and leaves GPT-5 nano flat. The Stargazer table has eight columns: surprise only, surprise plus the option statistics as regressors, and per baseline the surprise with the prediction as submitted, naively scaled, and adjusted. A final cell reports the competition's own delta_r_squared on the full quarter.
• scale_by_implied_move in examples.option_stats implements both rules, with tests.
• The bundled 2026Q3 sample now carries the item on its six events, so the parsing cells run offline; the regression cells need the live archive with the item and baseline predictions.
• README, data/README.md, the package docstring, and pointers in notebooks 01 and 03 updated.

Test plan

• uv run ruff check . && uv run ruff format --check .
• uv run mypy
• uv run pytest (108 passed)
• EM_API_KEY= uv run pytest --nbmake notebooks/ (all five notebooks in sample mode)
• Live execution of notebook 04 against the prod 2026Q3 archive; table and chart inspected

🤖 Generated with Claude Code