1[2 {"id":"C1","text":"The model has lower out-of-sample RMSE than seasonal naive at h=1, 2 and 4 (relative RMSE improvements about 0.19, 0.14, 0.17).","results_keys":["rmse_snaive_h1","rmse_model_h1","rmse_snaive_h2","rmse_model_h2","rmse_snaive_h4","rmse_model_h4","rel_improvement_rmse_h1","rel_improvement_rmse_h2","rel_improvement_rmse_h4"],"status":"supported","caveat":"Point estimates over 40 origins; no significance test at h=1,2; results driven partly by the 2021-23 surge."},3 {"id":"C2","text":"The model also has lower MAE than seasonal naive at all three horizons.","results_keys":["mae_snaive_h1","mae_model_h1","mae_snaive_h2","mae_model_h2","mae_snaive_h4","mae_model_h4"],"status":"supported","caveat":"Point estimates only."},4 {"id":"C3","text":"At h=4 the mean squared-error loss differential (2.57 sq. pp, favouring the model) is not statistically distinguishable from zero: HAC (Bartlett, lag 4) DM = 1.34, two-sided p = 0.18.","results_keys":["loss_diff_mean_h4","loss_diff_se_h4","hac_lags_h4","hac_kernel_h4","dm_stat_h4","dm_pvalue_h4"],"status":"supported","caveat":"Asymptotic normal approximation with 40 overlapping observations."},5 {"id":"C4","text":"The non-significance at 5% is robust to the HAC lag/kernel choices tried (Bartlett lag 3 and 6, rectangular lag 3, HLN correction), with p-values between 0.07 and 0.23.","results_keys":["sensitivity_dm_h4"],"status":"supported","caveat":"SE varies materially with the lag (1.43 to 2.16) because of negative higher-order autocorrelation in d_t."}6]7