1[2 {"id":"C1","text":"At h=1 the base-effect random-walk-with-drift model has a lower RMSE than the naive forecast (0.550 vs 0.997 pp; 44.8% improvement) and the DM test rejects equal accuracy at 5% (DM 2.43, 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; the normal approximation for DM is fragile in this sample."},3 {"id":"C2","text":"At h=3 the model has a lower RMSE than naive (1.142 vs 1.420; 19.6% improvement) but the difference is not statistically significant (DM 1.55, p 0.122; bootstrap interval includes zero).","results_keys":["rmse_model_h3","rmse_naive_h3","rel_improvement_rmse_h3","dm_stat_h3","dm_pvalue_h3","conclusion_code_h3","boot_rmse_diff_ci95_lo_h3","boot_rmse_diff_ci95_hi_h3"],"status":"supported","caveat":"Overlapping h=3 errors; two h=3 actuals come after a major level revision."},4 {"id":"C3","text":"Seasonal naive is the worst baseline at h=1 (RMSE 1.675) and AR(1) is close to naive at h=1 (0.941) but better at h=3 (1.187).","results_keys":["rmse_snaive_h1","rmse_ar1_h1","rmse_ar1_h3","rmse_snaive_h3"],"status":"supported","caveat":"Rankings based on 10 origins without formal tests among baselines."},5 {"id":"C4","text":"The first-release YoY growth of the h=1 target months differs on average by 0.548 pp from today's published YoY growth, the same order as the model's h=1 RMSE.","results_keys":["mean_abs_revision_h1_targets","rmse_model_h1"],"status":"supported","caveat":"Driven by a large level revision after mid-2026; not representative of typical revision size."},6 {"id":"C5","text":"Block-bootstrap 95% interval for RMSE_naive minus RMSE_model at h=1 excludes zero, and the model wins at 8 of 10 origins.","results_keys":["boot_rmse_diff_ci95_lo_h1","boot_rmse_diff_ci95_hi_h1","share_origins_model_smaller_sq_error_h1","random_seed"],"status":"supported","caveat":"Bootstrap with 10 observations is itself unreliable."}7]8