Reproduce “Conformal Prediction Regions for Time Series using Linear Complementarity Programming” on OpenResearch

About this paper

Matthew Cleaveland, Insup Lee, George J. Pappas, Lars Lindemann

This research introduces an optimization-based method utilizing Linear Complementarity Programming to construct less conservative yet statistically valid conformal prediction regions for time series. The approach consistently yields significantly tighter prediction regions, reducing their size compared to union bound methods, while maintaining desired coverage probabilities in dynamic system simulations.

Key insights

• Optimizing weights for a combined non-conformity score across a trajectory effectively mitigates the conservatism of union-bound approaches for time series conformal prediction.
• The complex MILCP optimization problem can be equivalently and more efficiently solved as an LCP, validating its computational tractability for real-world applications.
• The method provides statistically valid prediction regions for the entire trajectory with a specified confidence level, directly addressing limitations of asymptotic guarantees or overly conservative bounds.

Summary by alphaXiv.