Skip to content

Training-Free Probabilistic Time-Series Forecasting with Conformal Seasonal Pools

Home / Papers / Training-Free Probabilistic Time-Series Forecasting with Conformal Seasonal Pools

Training-Free Probabilistic Time-Series Forecasting with Conformal Seasonal Pools

Authors: Valery Manokhin Date: 2026-05-05 Paper ID: openalex:2605.03789

Summary

This paper introduces Conformal Seasonal Pools (CSP), a training-free, non-parametric probabilistic forecasting framework that enhances reliability in safety-critical applications. By combining empirical observations from matching seasonal periods with historical residuals, the method avoids the calibration pitfalls common in deep learning-based models, such as systematic under-coverage. Empirical results across six benchmark datasets show that CSP provides superior predictive intervals and computational efficiency compared to deep non-parametric alternatives. The author argues that such training-free conformal methods should serve as mandatory baselines for evaluating learned time-series models.

Key Contributions

  • Introduces Conformal Seasonal Pools (CSP), a training-free probabilistic forecasting framework that generates predictions by mixing historical seasonal empirical draws with signed residuals around a seasonal naive baseline.
  • Demonstrates significant performance improvements over DeepNPTS across six standard benchmarks (electricity, exchange_rate, solar_energy, taxi, traffic, wikipedia) in terms of CRPS, normalized mean quantile loss, and empirical coverage.
  • Achieves superior calibration (mean coverage 0.89 vs 0.66) while running 500x faster than deep learning baselines, highlighting the critical failure of parametric models in safety-sensitive forecasting tasks.

Open Questions & Future Work

Key Concepts

  • conformal-seasonal-pools: A training-free probabilistic forecasting method that generates prediction intervals by combining seasonal empirical observations with historical residuals.

Archivist Review

The paper introduces a highly effective training-free forecasting mechanism, Conformal Seasonal Pools, which challenges the assumption that deep learning models are necessary for high-quality probabilistic forecasting. I have approved this as a key concept for the vault. I also approved the open question regarding finite-sample coverage guarantees, as it addresses a fundamental theoretical boundary for applying conformal inference to serially dependent time-series data.

Approved Concepts

  • Conformal Seasonal Pools: Represents a new class of training-free, non-parametric conformal forecasting frameworks that serve as a strong performance and calibration benchmark for learned models.

Approved Open Questions

  • Finite-Sample Coverage Guarantees: This is a fundamental theoretical limitation for conformal prediction in time-series, separating practical empirical performance from rigorous safety requirements.

Metadata & Links

url
https://arxiv.org/abs/2605.03789
paper_id
2605.03789
paper_source
openalex
domain
time-series
tags
probabilistic-forecastingtime-series-forecastingconformal-predictioncalibration
architectures
datasets
concept_slugs
conformal-seasonal-pools
dataset_slugs
skill
TimeSeriesSkill
created_at
2026-05-08T05:43:23Z