Abstract
Conformal prediction offers a powerful framework for buildingdistribution-free prediction intervals for exchangeable data. Existing methodsthat extend conformal prediction to sequential data rely on fitting arelatively complex model to capture temporal dependencies. However, thesemethods can fail if the sample size is small and often require expensiveretraining when the underlying data distribution changes. To overcome theselimitations, we propose Reservoir Conformal Prediction (ResCP), a noveltraining-free conformal prediction method for time series. Our approachleverages the efficiency and representation learning capabilities of reservoircomputing to dynamically reweight conformity scores. In particular, we computesimilarity scores among reservoir states and use them to adaptively reweightthe observed residuals at each step. With this approach, ResCP enables us toaccount for local temporal dynamics when modeling the error distributionwithout compromising computational scalability. We prove that, under reasonableassumptions, ResCP achieves asymptotic conditional coverage, and we empiricallydemonstrate its effectiveness across diverse forecasting tasks.