Crime forecasts are sensitive to the spatial discretizations on which they are defined. Furthermore, while the Predictive Accuracy Index (PAI) is a common evaluation metric for crime forecasts, most crime forecasting methods are optimized using maximum likelihood or other smooth optimization techniques. Here we present a novel methodology that jointly (1) selects an optimal grid size and orientation and (2) learns a scoring function with the aim of directly maximizing PAI. Our method was one of the top performing submissions in the 2017 NIJ Crime Forecasting challenge, winning 9 of the 20 PAI categories under the name of team PASDA. We illustrate the model on data provided through the competition from the Portland Police Department.