# Contract (P0) — F3-conf-01

Question. At the end of month t, using only data published by then, how accurately can Canadian monthly real GDP YoY growth for t+1 and t+3 be forecast; does a model adding LFS employment YoY growth beat the naive forecast? Estimands. Over 48 expanding-window origins t = 2022-05 … 2026-04 and h ∈ {1,3}: RMSE and MAE (pp) of (a) naive = gdp_yoy(t-2), (b) one model; and 1 − RMSE_model/RMSE_naive per horizon. Meaningful as pseudo-real-time (publication-timing) evaluation on the current vintage; not real-time validity. 48 errors per horizon, overlapping/autocorrelated -> small-sample uncertainty. Information rules. At origin t: gdp_* usable only for months ≤ t−2; emp_* only for months ≤ t−1. Applies to estimation sample (training targets too), model selection, transforms and plugged-in values. Outputs. out/forecasts.csv (192 rows; origin_date,target_date,horizon,forecast,actual,model,last_target_date_used,last_predictor_date_used); out/results.json (task_id, model_name, model_description, n_origins, first_origin, last_origin, last_target_date_available, last_predictor_date_available, 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. Any use of gdp at > t−2 or emp at > t−1; missing targets for t+h (latest target 2026-07 = 2026-04+3, available); model chosen by looking at test errors.