Distributional conformal prediction.

Area of Science:

• Statistics
• Econometrics
• Machine Learning

Background:

• Constructing reliable prediction intervals is crucial for decision-making in various fields.
• Existing methods often struggle with heteroskedasticity and model misspecification.
• Accurate prediction intervals are essential for forecasting, synthetic controls, and causal inference.

Purpose of the Study:

• To propose a robust method for constructing conditionally valid prediction intervals.
• To extend the applicability of prediction intervals to diverse prediction problems.
• To ensure validity under various challenging conditions, including heteroskedasticity and model misspecification.

Main Methods:

• The method utilizes the probability integral transform and permuted ranks.
• It leverages models for conditional distributions, such as quantile and distribution regression.
• A "shape" adjustment is proposed for optimizing prediction intervals.

Main Results:

• The method constructs conditionally valid prediction intervals.
• It demonstrates approximate conditional validity under consistent estimation.
• Approximate unconditional validity is shown under model misspecification, overfitting, and with time series data.

Conclusions:

• The proposed method offers a robust approach to prediction interval construction.
• It is applicable to a wide range of prediction tasks, enhancing their reliability.
• The method provides a valuable tool for improving the accuracy of statistical predictions.