# Contract (P0)

Question. Over 36 rolling expanding-window origins (2023-07-01 … 2026-06-01), does a simple autoregressive-type model forecast the US unemployment rate (UNRATE, %, SA, monthly, final vintage) better than the naive no-change forecast at h=1 and h=3 months?

Estimands. Per horizon h∈{1,3}: out-of-sample RMSE and MAE (percentage points) of (a) naive, (b) one pre-specified model; relative improvement 1 − RMSE_model/RMSE_naive. Meaningful as descriptive forecast-accuracy measures on the revised (final-vintage) series over this specific 36-origin window; not real-time accuracy; small sample (≤36 pairs, overlapping for h=3).

Information rules. Forecast at origin t uses only observations dated ≤ t (estimation, selection, transformation). Naive = latest available observation ≤ t. Target t+h with missing actual stays in ledger with empty actual, excluded from errors; report n evaluated per horizon.

Outputs. out/forecasts.csv (144 rows: origin_date,target_date,horizon,forecast,actual,model); out/results.json (keys: task_id, model_name, model_description, n_origins, first_origin, last_origin, n_evaluated_h1, n_evaluated_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.

Invalidating conditions. Use of data dated > t at origin t; model chosen by looking at evaluation-period errors; missing actuals treated as zeros; fewer than 36 origins available in data.