1[2 {3 "id": "C1",4 "text": "The pre-specified ARIMA(2,1,0) model does not improve on the naive no-change forecast at h=1: its RMSE is about 2.3% higher.",5 "results_keys": [6 "rmse_model_h1",7 "rmse_naive_h1",8 "rel_improvement_rmse_h1",9 "mae_model_h1",10 "mae_naive_h1"11 ],12 "status": "supported",13 "caveat": "35 evaluated pairs; final-vintage data; only one model specification."14 },15 {16 "id": "C2",17 "text": "At h=3 the ARIMA(2,1,0) model is also slightly worse than naive: RMSE about 0.5% higher and MAE also higher.",18 "results_keys": [19 "rmse_model_h3",20 "rmse_naive_h3",21 "rel_improvement_rmse_h3",22 "mae_model_h3",23 "mae_naive_h3"24 ],25 "status": "supported",26 "caveat": "Overlapping 3-step errors; the difference is very small."27 },28 {29 "id": "C3",30 "text": "The h=1 loss difference is statistically significant by the Diebold-Mariano test (p about 0.02); the h=3 difference is not (p about 0.28).",31 "results_keys": [32 "dm_stat_h1",33 "dm_pvalue_h1",34 "dm_stat_h3",35 "dm_pvalue_h3"36 ],37 "status": "supported",38 "caveat": "The DM test is not designed for nested models and the forecasts are nearly identical, so the p-values are approximate."39 },40 {41 "id": "C4",42 "text": "35 of the 36 origin pairs per horizon were evaluated, because the 2025-10-01 observation is missing.",43 "results_keys": [44 "n_evaluated_h1",45 "n_evaluated_h3",46 "n_origins",47 "data_issues"48 ],49 "status": "supported",50 "caveat": ""51 },52 {53 "id": "C5",54 "text": "The estimated AR coefficients on monthly changes in the last window are close to zero, so the model's forecasts are nearly identical to naive.",55 "results_keys": [56 "ar1_coef_last_window",57 "ar2_coef_last_window"58 ],59 "status": "supported",60 "caveat": "Probably influenced by the 2020 COVID outlier kept in the sample."61 }62]