I would recommend proc glmselect, which is at the heart of the predictive regression models task in SAS Studio. The distribution of your dependent, response or target variable is a convolution of several processes that is likely to have some strange distribution that would not meet the usual gaussian/normal residual assumption required for ordinary least squares regression provided by proc reg. If you have enough (aggregated) observations to be able to afford a sampling split into training and validation data that would yield the most reliable inferences, otherwise use AICC as the selection criteria, and allow the store hierarchy to be split up for model selection.
... View more