New conformal prediction methods ensure nested sets across all coverage levels
Two novel online methods guarantee strictly nested prediction sets with stable coverage.
Conformal prediction has long offered finite-sample coverage guarantees, but most online extensions only handle a single coverage level. This creates a gap for real-world scenarios where different users need calibrated uncertainty estimates at varying risk tolerances simultaneously—think weather forecasts, macroeconomic predictions, or risk management dashboards.
Eduardo Ochoa Rivera and Ambuj Tewari address this with two novel online conformal prediction methods that produce strictly nested prediction sets across all coverage levels. By framing the problem as online optimization, they enforce monotonicity constraints with small regret, effectively controlling quantile estimation error. Empirical results on synthetic and real-world forecasting datasets show stable coverage across all levels and improved statistical efficiency compared to existing baselines.
- Two novel online conformal prediction methods output strictly nested prediction sets across a range of coverage levels.
- Approach uses online optimization to enforce monotonicity, achieving small regret and quantile estimation error control.
- Empirical results on forecasting tasks show stable coverage and improved efficiency over existing baselines.
Why It Matters
Enables calibrated uncertainty estimates for diverse risk tolerances in forecasting, risk management, and decision-making systems.