Risk Bounds and Calibration for a Smart Predict-then-Optimize Method