Tube Loss: A New Method for Prediction Interval Estimation
A research paper on arXiv proposes 'Tube Loss', a novel loss function for estimating prediction intervals (PIs) in regression. The method yields intervals that asymptotically achieve a prespecified confidence level t ∈ (0,1), with a theoretical proof provided. Users can adjust a parameter to shift the interval, capturing denser regions of the response variable's probability distribution and sharpening interval width, especially useful for skewed conditional distributions. The approach also allows trade-offs between coverage and width. The paper is available at https://arxiv.org/abs/2412.06853.
Key facts
- Tube Loss is a novel loss function for prediction interval estimation.
- It asymptotically attains prespecified confidence level t ∈ (0,1).
- A theoretical proof of asymptotic confidence level attainment is given.
- Users can control a parameter to move the interval up or down.
- This helps capture denser regions of the probability distribution.
- It is especially useful for skewed conditional distributions.
- The method can trade-off between coverage and interval width.
- The paper is available on arXiv with ID 2412.06853.
Entities
Institutions
- arXiv