Introduction
In high-stakes applications such as medical diagnostics or autonomous navigation, uncertainty estimation is as critical as point prediction. Quantile regression is the standard "vanilla" approach for generating prediction intervals. However, a persistent issue in practice is "under-coverage": a model tasked with providing 90% coverage often fails to meet this target, producing intervals that are systematically too narrow. This paper investigates whether this phenomenon is a result of data quality or an inherent mathematical bias in high-dimensional settings.
Theoretical Framework
The authors analyze linear quantile regression through the lens of the pinball loss function within a "high-dimensional proportional regime." In this setting, the number of features ($d$) is large and grows proportionally with the number of training samples ($n$). By applying the Convex Gaussian Comparison Theorem (CGMT), the researchers characterize the behavior of Empirical Risk Minimization (ERM) solutions as $n, d \to \infty$. This approach allows for a precise disentanglement of error sources, specifically identifying how inaccuracies in the learned bias term and feature weights ($w$) impact final coverage.
The Under-Coverage Phenomenon
The study reveals that under-coverage is not a fluke but a predictable consequence of model training in finite-sample regimes. As the ratio of features to data increases, the gap between the nominal target coverage ($\alpha$) and the actual realized coverage widens.
The primary driver of this bias is the estimation of feature weights. While errors in the intercept can fluctuate, errors in the weights consistently push the model toward overconfidence, resulting in intervals that are too tight for the underlying data distribution.
Quantifying the Bias: The $d/n$ Rule
A key contribution of this work is a simple, actionable rule of thumb to estimate actual coverage: $$\text{Actual Coverage} \approx \alpha - (\alpha - 1/2) \cdot \frac{d}{n}$$ For example, if a practitioner uses 100 features with 1,000 samples ($d/n = 0.1$) to aim for 90% coverage ($\alpha = 0.9$), the actual coverage will likely be around 86%.
This linear relationship holds remarkably well across various settings. In the extreme over-parameterized regime ($d > n$), the model's capacity to overfit the uncertainty becomes so great that it may default to a baseline 50% coverage, regardless of the requested $\alpha$.
Empirical Validation
The theoretical findings were validated across synthetic Gaussian models and several real-world datasets, including the Medical Expenditure Panel Survey (MEPS) and the Community and Crimes dataset.
- Neural Networks: The bias is exacerbated in high-capacity models like Multi-Layer Perceptrons (MLPs). On the "Community" dataset, an MLP aiming for 90% coverage achieved only 63.09%.
- Model Specification: By using "clean" labels from a known linear model, the authors demonstrated that the bias persists even when the model is perfectly specified, confirming it is an inherent property of the optimization process rather than "bad data."
Conclusion and Future Directions
This research highlights a fundamental limitation of standard quantile regression. The systematic under-coverage observed in high-dimensional models necessitates the use of post-processing techniques, such as Conformal Prediction, to ensure reliable uncertainty estimates. Future work may explore "de-biasing" the pinball loss function directly during the training phase to mitigate these effects without requiring post-hoc corrections.
Summary updated