# Contract (P0) — T1-forecast

**Question.** In real time (only data published at each of 10 month-end origins, 2025-10-31 … 2026-07-31), how accurately can US retail & food services sales YoY growth g_m = 100·ln(L_m/L_{m-12}) be forecast h = 1 and h = 3 months beyond the last published month last(V), and does my own model beat the naive (no-change) forecast when errors are measured against the first-release YoY of the target month?

**Estimands** (h ∈ {1,3}, n = 10 origins each):
(a) RMSE and MAE of naive, seasonal naive, AR(1) and my model vs `yoy_first_release`;
(b) 1 − RMSE_model/RMSE_naive;
(c) DM statistic (d_t = e²_naive − e²_model), Bartlett/Newey–West LRV with lag h−1, no HLN correction, two-sided normal p-value;
(d) mean over the 10 h=1 target months of |g^current_T − yoy_first_release_T|.
Meaningful under: the forecast-error loss differential is covariance-stationary (DM), first release is the agreed "truth"; with n = 10 inference is very weak (DM asymptotics unreliable).

**Outputs.** out/forecasts.csv (80 rows; origin_date,target_date,horizon,forecast,actual_first_release,model,vintage_used), out/results.json (keys listed in TASK.md: task_id, model_name, model_description, n_origins, first_origin, last_origin, rmse/mae_{naive,snaive,ar1,model}_h{1,3}, rel_improvement_rmse_h{1,3}, dm_stat/pvalue_h{1,3}, conclusion_code_h{1,3}, mean_abs_revision_h1_targets, feasibility, feasibility_reason, data_issues, n_input_rows, n_analysed, random_seed), out/table1.csv (model,horizon,rmse,mae,n), out/figure1.png, out/report.md, out/analysis.py, out/claims.json; plus protocol files in out/state/ and out/exclusions.csv.

**Information rules.** At origin V: only RSAFS vintage V, PAYEMS vintage V, UMCSENT months ≤ month(V). Target = last(V)+h where last(V) = last month in RSAFS vintage V (publication delays exist). RSAFS_current.csv only for (d). first_release.csv only as actuals (never as forecast input).

**Invalidating conditions.** Missing target in first_release.csv; vintage lacking needed history; any use of current data / future vintages in forecasts; model tuned on the evaluation origins' outcomes (avoided: model pre-specified before evaluation); negative LRV estimate in DM (would make statistic undefined → null).