# Contract (P0)

Question: Over 36 expanding-window origins (2023-07-01..2026-06-01), does a simple AR-type model beat the naive no-change forecast of US UNRATE at h=1 and h=3 months? Estimands: out-of-sample RMSE and MAE (pp) of naive and model per horizon, over (origin,h) pairs with non-missing actual; rel. improvement 1 - RMSE_model/RMSE_naive. Meaningful as forecastability of the final-vintage (revised) series; not real-time accuracy. 36 origins -> small-sample, noisy. Outputs: out/forecasts.csv (144 rows: origin_date,target_date,horizon,forecast,actual,model), out/results.json (keys per TASK.md table), out/report.md (6 sections), out/analysis.py (deterministic), out/claims.json. Information rules: forecast at t uses only obs dated <= t (estimation, selection, transformations). Naive = latest available obs <= t. Missing actual -> row kept with empty actual, excluded from stats; report n evaluated per horizon. No real-time claims. Invalidity: using obs > t; tuning model spec on the evaluation errors; missing data making estimation impossible (not the case).