Retailers increasingly apply price markdowns for their seasonal products. Efficiency of these markdown applications is driven by the accuracy of empirical models, especially toward the end of a selling season. In the literature, recent sales are recognized to be more important than older sales data for estimating the current period’s demand for a given markdown level. The importance difference between the weeks of a selling season is addressed by weighted least squares (WLS) method with continuous weight functions of time. This study suggests a generalization of the weight functions and a method for optimizing their shape and discretization parameters to stimulate the predictive accuracy of models. We find that addressing the importance difference of recent sales observations using our generalized weight functions improves the forecast accuracy by up to 20%, and most of the improvement stems from our weight discretization method.