Simulating classification models to evaluate Predict-Then-Optimize methods