Calibrated AI uncertainty inside robust optimization
Distributionally robust optimization is one of the fastest-growing methods in operations research, and uncertainty estimation is rising in the AI journals, yet only two papers in the archive touch both. Can calibrated uncertainty from machine learning decide how much robustness an optimization model buys?
What prompted this
On the trends page, "distributionally robust optimization" is growing in the operations research journals (22 papers in July–September 2026 against about 15 expected) and "uncertainty estimation" in the AI journals (39 against 23). Since October 2025, 195 papers carry a robust-optimization keyword (158 in operations research), but only 2 of them also carry a conformal-prediction, uncertainty-quantification, uncertainty-estimation or calibration keyword. Uncertainty quantification (105 papers) and uncertainty estimation (117) stay almost entirely in the AI journals, while "demand uncertainty" is the second most frequent operations research keyword of the year.
Questions
- If the size of the ambiguity set (say, a Wasserstein radius) comes from conformal prediction intervals of a demand forecast, does it beat a cross-validated radius out of sample?
- How much cost does too much robustness waste in inventory problems such as the newsvendor model?
- What happens under distribution shift, when the forecast's calibration breaks?
Data that could help
Public retail demand data such as the M5 competition data, and standard inventory and network-design instances.
Next step
Solve a newsvendor problem on M5 data with conformal intervals and with a cross-validated Wasserstein radius, and compare cost and service level.