Diebold–Mariano on two loss series of the same forecasts' targets, in
order: d_t = lossA_t − lossB_t, its mean over a HAC variance with
horizon − 1 lags (h-step forecasts' errors overlap that far), times
HLN's √((n + 1 − 2h + h(h − 1)/n)/n), against t(n − 1). For forecasts made
every step origins over horizon-long windows, the windows overlap for
⌈horizon/step⌉ origins: pass that as horizon.
Diebold–Mariano on two loss series of the same forecasts' targets, in order: d_t = lossA_t − lossB_t, its mean over a HAC variance with
horizon− 1 lags (h-step forecasts' errors overlap that far), times HLN's √((n + 1 − 2h + h(h − 1)/n)/n), against t(n − 1). For forecasts made everysteporigins overhorizon-long windows, the windows overlap for ⌈horizon/step⌉ origins: pass that ashorizon.