Dissecting the Impact of Model Misspecification in Data-driven Optimization
The paper compares two methods for data-driven optimization: traditional estimate-then-optimize (ETO) and integrated estimation-optimization (IEO). It shows that IEO provides a 'universal double benefit' on the top two regret terms when the model is misspecified, meaning it performs significantly better in those critical aspects. However, ETO can be advantageous when the model is nearly correct.