1[2 {"id":"C1","text":"Real-time AR(1) (1985+, pandemic pairs excluded) achieved RMSE 1.48 pp at h=1 vs 2.01 pp for naive, a 26% reduction, against first releases over 8 origins.","results_keys":["rmse_model_h1","rmse_naive_h1","rel_improvement_rmse_h1"],"status":"supported","caveat":"Not statistically significant (HLN-DM p=0.19, T=8); driven largely by two origins."},3 {"id":"C2","text":"At h=2 the AR(1) RMSE was 1.40 pp vs 2.27 pp for naive, a 38% reduction, and the model was better at all 8 origins.","results_keys":["rmse_model_h2","rmse_naive_h2","rel_improvement_rmse_h2"],"status":"supported","caveat":"HLN-DM p=0.13; overlapping targets make origins dependent."},4 {"id":"C3","text":"MAE: naive 1.69/1.81 pp and AR(1) 1.22/1.02 pp at h=1/h=2.","results_keys":["mae_naive_h1","mae_naive_h2","mae_model_h1","mae_model_h2"],"status":"supported","caveat":"8 observations per horizon."},5 {"id":"C4","text":"First releases of the h=1 target quarters were later revised by 0.56 pp on average in absolute value (current vintage 2026-10-30), a sizable fraction of forecast RMSE.","results_keys":["mean_abs_revision_h1_targets"],"status":"supported","caveat":"Recent quarters (2025Q4) have had fewer revision rounds than older ones."},6 {"id":"C5","text":"A simple AR model significantly beats the naive forecast in real time.","results_keys":["rel_improvement_rmse_h1","rel_improvement_rmse_h2"],"status":"unsupported","caveat":"Point estimates favour the model but the 8-origin sample does not yield statistical significance."}7]8