HANDOFF.md

Handoff

Read the Space's About and Data & submission tabs once.

The task

• Stand at the end of 2018. Forecast 2019–2025 for 210 cities × 4 metrics = 5,880 rows in submission.csv (the rows are pre-filled; we fill predicted_value).
• Metrics: annual_avg_temp (Avg Temp), total_precip_in (ToT Prcp), hottest_day_temp (Max Tmax), cdd (cdd), all from annual_indices.
• Score = MAE per metric relative to the 30-year-average baseline, averaged. Live score covers 2019–21, final score covers 2022–25.
• Organizers say most of the gain is in temp and CDD. For precip and hottest day, matching the 30-year average is already fine.

Data rule (can disqualify us)

• Forecast inputs may only be weather up to 2018-12-31, plus non-weather-driven data from any year (census, land cover, tree canopy, elevation).
• No 2019+ weather or anything driven by it (streamflow, air quality, energy demand). No values from LLMs or pretrained models that saw post-2018 data.
• 2019+ data is fine for the policy idea.
• If you use Claude/ChatGPT/Copilot, write down which tool, what for, and what you checked yourself. It goes in the README.

Already in the repo

• code/baseline.py runs the 3 official baselines, validates them (train ≤2010, predict 2011–18), picks the best per metric, and writes a valid submission.csv. Run this first so we always have something submittable.
• code/check_submission.py (from the template) runs the same checks as the Space.

Your tasks

1. python code/baseline.py, then python code/check_submission.py. Log anything that breaks in friction_log.md.
2. Improve temp + CDD. Ideas: trend fitted on a shorter or smoothed window, pooling trends across nearby cities, a model on lag features + city lat/lon. Always compare against the validation MAE from the baseline.
3. Interoperability (20%): add one external dataset that's allowed as an input (e.g. census population/density, NLCD land cover, tree canopy, elevation) and join it to city_info. Save the processed version in data/ and describe it in data/README.md (source, license, how prepared).
4. Make the final model one script that reproduces submission.csv from data/.
5. Leave me short notes for the README: method in a few lines, validation MAE table, 1 figure.

Rules

• random_state=42 everywhere.
• Never commit tokens.
• Log friction as you go.