# Contract (P0)

Question. In real time (10 month-end ALFRED origins 2025-10-31..2026-07-31), how accurately can YoY growth of US retail & food services sales (RSAFS), g_m = 100 ln(L_m/L_{m-12}), be forecast h=1 and h=3 months beyond last(V), the last month in vintage V; does an own model beat the naive (no-change) forecast, evaluated against first-release YoY (data/first_release.csv)?

Estimands. (a) RMSE, MAE per horizon for naive, seasonal naive, AR(1) (OLS on g_m with 1993-02<=m<=last(V), iterated) and own model; (b) 1 - RMSE_model/RMSE_naive; (c) DM stat model vs naive, d_t = e2_naive - e2_model, Bartlett LRV with lag h-1, no HLN correction, two-sided normal p; (d) mean |g_current_T - yoy_first_release_T| over the 10 h=1 targets. Meaningful under: first release as the "truth"; n=10 so DM asymptotics are only indicative; overlapping h=3 targets.

Outputs. out/forecasts.csv (80 rows), out/results.json (keys per TASK.md table), out/table1.csv, out/figure1.png, out/report.md (6 sections), out/analysis.py, out/claims.json, plus state files and out/exclusions.csv.

Information rules. Forecast at V uses only RSAFS vintage V, PAYEMS vintage V, UMCSENT months <= month(V). RSAFS_current only for (d). Target = last(V)+h. vintage_used <= origin < target.

Invalidity. Using current data or later vintages in forecasts; tuning the model on evaluation actuals; missing first-release actuals for targets (would make task partial).