The idea
Two forecasts can have similar errors yet lead to very different actions. Elmachtoub and Grigas study learning in the context of a downstream optimisation problem, measuring prediction errors by the decisions they induce.
What the authors established
The paper develops the SPO loss and a convex surrogate, SPO+, with consistency results under stated conditions. Shortest-path and portfolio experiments illustrate how using the decision structure during training can improve the resulting decisions.
Applications
Capacity planning and purchasing connect forecast errors to financial and service outcomes. Evaluating decisions on held-out scenarios reveals costs that forecast accuracy alone can miss.
Where care is needed
The optimisation model must capture the relevant objective and constraints. The paper’s guarantees have assumptions and do not turn every prediction task into a tractable training problem.
Sources and credits
- Smart “Predict, then Optimize”
Adam N. Elmachtoub, Paul Grigas, published
Management Science (2022); first published online in 2021.
The SPO framework and SPO+ loss are the authors’ contributions. We explain the distinction between predictive error and decision quality.