[
 {"id": "C1", "text": "The base-effect drift model has lower RMSE than the naive forecast at h=1 (0.550 vs 0.997 pp; 44.8% improvement) and the Diebold-Mariano test rejects equal accuracy at 5% (p=0.015).", "results_keys": ["rmse_model_h1", "rmse_naive_h1", "rel_improvement_rmse_h1", "dm_stat_h1", "dm_pvalue_h1", "conclusion_code_h1"], "status": "supported", "caveat": "Only 10 origins; HLN-corrected p=0.047 is borderline."},
 {"id": "C2", "text": "At h=3 the model improves RMSE over naive by 19.6% (1.142 vs 1.420 pp), but the difference is not statistically significant (DM p=0.12).", "results_keys": ["rmse_model_h3", "rmse_naive_h3", "rel_improvement_rmse_h3", "dm_stat_h3", "dm_pvalue_h3", "conclusion_code_h3"], "status": "supported", "caveat": "Overlapping 3-step errors and n=10 make the long-run variance estimate noisy."},
 {"id": "C3", "text": "The model has the lowest RMSE and MAE of all four forecasts (naive, seasonal naive, AR(1), model) at both horizons.", "results_keys": ["rmse_model_h1", "rmse_ar1_h1", "rmse_snaive_h1", "rmse_model_h3", "rmse_ar1_h3", "rmse_snaive_h3", "mae_model_h1", "mae_model_h3"], "status": "supported", "caveat": "Differences versus AR(1) at h=3 are small (1.142 vs 1.187) and untested."},
 {"id": "C4", "text": "The mean absolute revision between first-release and current YoY growth for the h=1 target months is 0.548 pp, about the size of the best h=1 forecast RMSE.", "results_keys": ["mean_abs_revision_h1_targets", "rmse_model_h1"], "status": "supported", "caveat": "Current vintage includes a large benchmark level revision; future revisions may change this."},
 {"id": "C5", "text": "PAYEMS and consumer sentiment do not improve real-time forecasts of retail-sales growth.", "results_keys": [], "status": "not_evaluated", "caveat": "These predictors were not tested in this project."}
]