1[2 {"id":"C1","text":"The employment-augmented direct model dar_emp has lower out-of-sample RMSE than the naive forecast at h=1 (0.743 vs 0.814 pp; relative improvement 0.087).","results_keys":["rmse_model_h1","rmse_naive_h1","rel_improvement_rmse_h1"],"status":"supported","caveat":"Current-vintage data; 48 origins; size of gain sensitive to COVID exclusion (sens_nocovidexcl_rel_improvement_rmse_h1=0.027)."},3 {"id":"C2","text":"At h=3 the model lowers RMSE relative to naive by 0.224 (0.885 vs 1.140 pp).","results_keys":["rmse_model_h3","rmse_naive_h3","rel_improvement_rmse_h3"],"status":"supported","caveat":"DM test significant only at 10% (p=0.093); evaluation period is post-pandemic disinflationary growth."},4 {"id":"C3","text":"The model also has lower MAE than naive at both horizons.","results_keys":["mae_model_h1","mae_naive_h1","mae_model_h3","mae_naive_h3"],"status":"supported","caveat":"Same sample caveats as C1-C2."},5 {"id":"C4","text":"The h=1 improvement is statistically significant at 5% by a HLN-corrected Diebold-Mariano test; the h=3 improvement only at 10%.","results_keys":["dm_stat_h1","dm_pvalue_h1","dm_stat_h3","dm_pvalue_h3"],"status":"supported","caveat":"Small sample (48), overlapping errors at h=3; asymptotic approximation."},6 {"id":"C5","text":"Excluding COVID-window training pairs changes the magnitude but not the sign of the improvement.","results_keys":["sens_nocovidexcl_rel_improvement_rmse_h1","sens_nocovidexcl_rel_improvement_rmse_h3"],"status":"supported","caveat":"Only one alternative treatment examined."},7 {"id":"C6","text":"All forecasts respect the publication-lag rules (GDP <= t-2, employment <= t-1); last values available at the final origin are 2026-02 and 2026-03.","results_keys":["last_target_date_available","last_predictor_date_available","n_analysed"],"status":"supported","caveat":"Timing simulation only; no real-time vintage validity."}8]9