Adaptive Conformal Inference by Betting


{getToc} $title={Table of Contents}

Summary

The paper proposes a method for adaptive conformal inference using betting, which achieves long-term coverage guarantee without requiring parameter tuning.

Highlights

  • The method uses parameter-free online convex optimization techniques, specifically the Krichevsky-Trofimov (KT) estimator and the Online Newton Step (ONS) method.
  • The approach is based on framing the task of constructing adaptive conformal predictors as a problem of learning the (1-α)-quantile of the nonconformity scores in an online fashion.
  • The method is shown to be effective in simulations and real-world datasets, including changepoint settings and stock price forecasting.
  • The approach is compared to existing methods, including online gradient descent (OGD) and scale-free online gradient descent (SF-OGD).
  • The method is shown to be robust to distribution shifts and can handle multi-step forecasting.
  • The approach is simple to implement and does not require explicit knowledge of the bound on the nonconformity scores.

Key Insights

  • The proposed method achieves long-term coverage guarantee without requiring parameter tuning, making it a practical and accessible choice for applications.
  • The use of parameter-free online convex optimization techniques allows the method to adapt to changes in the data distribution without requiring explicit knowledge of the bound on the nonconformity scores.
  • The method's ability to handle multi-step forecasting makes it suitable for applications where forecasting is done over a horizon, such as in finance and economics.
  • The comparison to existing methods shows that the proposed approach is competitive and can outperform other methods in certain scenarios.
  • The method's robustness to distribution shifts makes it suitable for applications where the data distribution is expected to change over time.
  • The simplicity of the method makes it easy to implement and integrate into existing workflows.
  • The method's ability to provide uncertainty estimates makes it suitable for applications where uncertainty quantification is important, such as in finance and healthcare.



Mindmap


Citation

Podkopaev, A., Xu, D., & Lee, K.-C. (2024). Adaptive Conformal Inference by Betting (Version 1). arXiv. https://doi.org/10.48550/ARXIV.2412.19318

Previous Post Next Post

Contact Form