1[2 {"id":"C1","text":"In real time over 8 origins, the naive forecast had RMSE 2.010 (h=1) and 2.265 (h=2) pp against first releases.","results_keys":["rmse_naive_h1","rmse_naive_h2","mae_naive_h1","mae_naive_h2"],"status":"supported","caveat":"Only 8 errors per horizon."},3 {"id":"C2","text":"The pre-specified AR(1) (1985Q1+, pandemic quarters masked) had lower RMSE than naive: 1.484 at h=1 and 1.398 at h=2, i.e. relative improvements of 26% and 38%.","results_keys":["rmse_model_h1","rmse_model_h2","rel_improvement_rmse_h1","rel_improvement_rmse_h2","mae_model_h1","mae_model_h2"],"status":"supported","caveat":"Point estimates in a short sample."},4 {"id":"C3","text":"The AR(1) advantage is not statistically significant at conventional levels (HLN-corrected DM p = 0.18 at h=1, 0.11 at h=2).","results_keys":["supp_dm_hln_stat_h1","supp_dm_hln_pvalue_two_sided_h1","supp_dm_hln_stat_h2","supp_dm_hln_pvalue_two_sided_h2"],"status":"supported","caveat":"DM test approximations are weak with n=8."},5 {"id":"C4","text":"First releases of the h=1 target quarters were later revised by 0.5625 pp on average in absolute value (to the 2026-10-30 vintage).","results_keys":["mean_abs_revision_h1_targets"],"status":"supported","caveat":"Later vintages may revise further."},6 {"id":"C5","text":"Both models' mean errors are small and negative (forecasts slightly too high on average).","results_keys":["supp_mean_error_naive_h1","supp_mean_error_naive_h2","supp_mean_error_model_h1","supp_mean_error_model_h2"],"status":"supported","caveat":"No bias test performed; 8 observations."}7]8