1[2 {"id":"C1","text":"The seasonal AR model has lower out-of-sample RMSE and MAE than the seasonal-naive forecast at h = 1, 2 and 4 over the 40 origins.","results_keys":["rmse_snaive_h1","rmse_model_h1","rmse_snaive_h2","rmse_model_h2","rmse_snaive_h4","rmse_model_h4","mae_snaive_h1","mae_model_h1","mae_snaive_h2","mae_model_h2","mae_snaive_h4","mae_model_h4"],"status":"supported","caveat":"Point estimates from 40 forecasts per horizon; no formal test at h=1,2."},3 {"id":"C2","text":"Relative RMSE improvements over seasonal-naive are about 0.23, 0.14 and 0.19 at h = 1, 2, 4.","results_keys":["rel_improvement_rmse_h1","rel_improvement_rmse_h2","rel_improvement_rmse_h4"],"status":"supported","caveat":"Sample includes the 2020-2023 shocks that dominate squared errors."},4 {"id":"C3","text":"At h = 4 the mean squared-error loss differential is positive (model better) but not statistically distinguishable from zero using a Bartlett HAC SE with lag 3 (p about 0.17).","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":"Result is bandwidth-sensitive: longer Bartlett lags (6, 8) give p below 0.05 because d_t has negative autocorrelation at lags 3-4; HLN correction gives p about 0.22."},5 {"id":"C4","text":"The h = 4 significance verdict depends on the HAC truncation lag.","results_keys":["sensitivity_h4"],"status":"supported","caveat":"Sensitivity analysis only; primary lag pre-specified as h-1=3."}6]7