 Asymmetric absolute-error loss

This loss assigns different positive costs to underestimation and overestimation of a scalar parameter. Here $\gamma$ penalizes estimates below the parameter and $\delta$ estimates above it. Equal costs give a multiple of absolute-error loss.