# 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 of one model per horizon; rel improvement 1 - RMSE_model/RMSE_naive. Meaningful as descriptive accuracy on the final-vintage series over this window (not real-time; small sample, n<=36). Outputs: out/forecasts.csv (144 rows: origin_date,target_date,horizon,forecast,actual,model); out/results.json (task_id, model_name, model_description, n_origins, first_origin, last_origin, n_evaluated_h1/h3, rmse/mae naive & model h1/h3, rel_improvement_rmse_h1/h3, feasibility, feasibility_reason, data_issues, n_input_rows, n_analysed, random_seed); out/report.md (Question; Data and exclusions; Method; Results; Uncertainty; Limitations and open issues); out/analysis.py; out/claims.json. Information rules: forecast at t uses only obs dated <= t, incl. model selection, estimation, transforms. Naive = latest obs <= t. Missing actual -> row kept with empty actual, excluded from stats. Invalid if: using data > t; model choice tuned on evaluation window outcomes; fabricated values. Impossible if data missing for all targets (not the case).