Search arXiv⌕ Search
Home/Archive/arXiv

arXiv · 2605.12679

Measures of predictive accuracy, miscalibration and discrimination

Łukasz Delong·Mario Wüthrich
View original source ↗Read source document ↗Export metadata ↓

Abstract

We study the evaluation of real-valued point predictors under the decision-theoretic framework of mean-consistent loss functions given by the Bregman divergences. We first derive a new version of Murphy's decomposition of the expected loss which does not directly include the response itself but only its predictors. We then relate the miscalibration and the discrimination component of Murphy's decomposition to Lorenz-curve-based accuracy measures that are widely used in practice. Besides the usual area between the concentration and Lorenz curves, ABC, we introduce a mean-squared version ABC$^2$ that mitigates some of the weaknesses of the original ABC in identifying mean-calibration. More importantly, both ABC and ABC$^2$ are shown to rely on predictor-dependent weights, so they fail to align with the class of mean-consistent scoring functions. In the same spirit, we derive a similar result for the widely used Gini score. These results indicate that ABC, ABC$^2$ and Gini scores may lead to misleading evaluation of point predictions when used for model selection; this gives support to use mean-consistent loss functions and their Murphy's decomposition. Finally, we study forecast dominance when Lorenz curves intersect. In the one-crossing case, we establish weaker dominance criteria for subclasses of Bregman divergences through third-degree stochastic dominance.

Explore related subjects

stat.ME

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Łukasz Delong, Mario Wüthrich. 2026-07-10. Measures of predictive accuracy, miscalibration and discrimination. https://arxiv.org/abs/2605.12679

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

Switchback Experiments under Geometric Mixing

The switchback is an experimental design that measures treatment effects by repeatedly turning an intervention on and off for a whole system. Switchback experiments are a robust way to overcome cross-unit spillover effects; however, they are vulnerable to bias from temporal carryovers. In this paper, we consider properties of switchback experiments in Markovian systems that mix at a geometric rate. We find that, in this setting, standard switchback designs suffer considerably from carryover bias: Their estimation error decays as $T^{-1/3}$ in terms of the experiment horizon $T$, whereas in the absence of carryovers a faster rate of $T^{-1/2}$ would have been possible. We also show, however, that judicious use of burn-in periods can considerably improve the situation, and enables errors that decay as $\log(T)^{1/2}T^{-1/2}$. Our formal results are mirrored in an empirical evaluation.

stat.ME↗
arXiv2026

Saddlepoint approximations for plug-in resampling

Resampling-based procedures can improve on normal approximations in sparse, large-scale testing problems, but their computational cost can be prohibitive. We recognize that several existing procedures belong to a faster plug-in resampling subclass, fixing fitted nuisance parameters during resampling. When the resampled statistic is a sum of conditionally independent terms, the saddlepoint approximation (SPA) for the resampling $p$-value offers further acceleration, replacing resampling with an analytical tail approximation. However, standard Edgeworth-based approximation-error bounds impose regularity conditions that are hard to verify for plug-in resampling laws. We use an alternative approach to establish a finite-sample relative-error bound for the Lugannani-Rice approximation under more tractable conditions, which we apply in two contexts. In statistical genetics, we identify response resampling procedures as the targets of existing SPAs and establish guarantees in a representative setting. In conditional independence testing, we introduce spaCRT, an SPA for the distilled conditional randomization test (dCRT), which has been applied successfully in biology. Our rates quantify the effects of sparsity and signal strength, with matching lower bounds in special cases. We additionally establish asymptotic Type-I error control of the corresponding plug-in resampling procedures under growing sparsity. In simulations and single-cell CRISPR data analysis, spaCRT closely approximates dCRT $p$-values and preserves its statistical performance while accelerating computation by up to 250-fold.

stat.ME↗
arXiv2026

NIRVAR: Network Informed Restricted Vector Autoregression

High-dimensional panels of time series often arise in finance and macroeconomics, where co-movements within groups of panel components occur. Extracting these groupings from the data provides a coarse-grained description of the complex system in question and can inform subsequent prediction tasks. We develop a novel methodology to model such a panel as a restricted vector autoregressive process, where the coefficient matrix is the weighted adjacency matrix of a stochastic block model. This network time series model, which we call the Network Informed Restricted Vector Autoregression (NIRVAR) model, yields a coefficient matrix that has a sparse block-diagonal structure. We propose an estimation procedure that embeds each panel component in a low-dimensional latent space and clusters the embedded points to recover the blocks of the coefficient matrix. Crucially, the method allows for network-based time series modelling when the underlying network is unobserved. We derive the bias, consistency and asymptotic normality of the NIRVAR estimator. Simulation studies suggest that the NIRVAR estimated embedded points are Gaussian distributed around the ground truth latent positions. On three applications to finance, macroeconomics, and transportation systems, NIRVAR outperforms competing models in terms of prediction and provides interpretable results regarding group recovery.

stat.ME↗