Article
Statistics & Probability
Torsten Hothorn, Achim Zeileis
Summary: This article discusses regression models for supervised learning problems with continuous response, suggesting a more general understanding of regression models as models for conditional distributions. Quantile regression forests are highlighted among algorithms estimating conditional distributions. A novel approach based on a parametric family of distributions characterized by their transformation function is proposed, along with a dedicated transformation tree algorithm for detecting distributional changes. Prediction intervals and inference procedures are provided by the resulting predictive distributions, making them fully parametric yet very general.
JOURNAL OF COMPUTATIONAL AND GRAPHICAL STATISTICS
(2021)
Article
Energy & Fuels
Norman Maswanganyi, Caston Sigauke, Edmore Ranganai
Summary: Predicting extreme electricity demand using different regression models and combining estimated quantiles at various levels can provide insights for decision-makers in power utilities. Comparing the models using proper scoring rules helps to determine the most accurate approach for predicting future electricity demand patterns.
Article
Mathematics, Applied
Yunyi Zhang, Dimitris N. Politis
Summary: In this paper, we propose a new bootstrap algorithm that generates a prediction interval controlling both the conditional coverage probability and the possibility of conditional under-coverage.
INFORMATION AND INFERENCE-A JOURNAL OF THE IMA
(2023)
Article
Automation & Control Systems
Maxime Cauchois, Suyash Gupta, John C. Duchi
Summary: The study introduces a novel conformal prediction method to construct valid predictive confidence sets in multiclass and multilabel problems, ensuring correct coverage and asymptotically consistent conditional coverage. To address the potential challenge of exponentially large confidence sets in multilabel prediction, tree-structured classifiers are built to efficiently account for interactions between labels.
JOURNAL OF MACHINE LEARNING RESEARCH
(2021)
Article
Mathematics, Applied
Zheng-Xin Wang, Yue-Qi Jv
Summary: A novel grey prediction model based on quantile regression technology (QGM(1,1) model) is proposed to address the issue of traditional grey prediction models being influenced by outliers and lacking stability. The QGM(1,1) model accurately describes the impact of independent variables on the range and shape of dependent variables, as well as captures tail characteristics of the distribution. Results show that the model significantly improves prediction accuracy and enhances robustness.
COMMUNICATIONS IN NONLINEAR SCIENCE AND NUMERICAL SIMULATION
(2021)
Article
Economics
Christopher Oconnor
Summary: Standard methodologies used to identify vulnerable households rely on distributional assumptions that may lead to classification errors. This paper demonstrates that quantile models can improve this identification by relaxing these assumptions. Quantile models are robust and easy to implement, making them suitable for policymakers. Applying this strategy to data from Uganda shows that it more accurately identifies the future poor compared to standard approaches. The study highlights the benefits of relaxing distributional assumptions when identifying vulnerable populations.
ECONOMIC MODELLING
(2023)
Article
Statistics & Probability
Qianqian Zhu, Songhua Tan, Yao Zheng, Guodong Li
Summary: This article proposes a novel conditional heteroscedastic time series model using the framework of quantile regression processes in the ARCH(& INFIN;) form of the GARCH model. The model can provide varying structures for conditional quantiles of the time series across different quantile levels, including the commonly used GARCH model as a special case. The article introduces a self-weighted composite quantile regression estimator to remedy the accuracy deterioration at high quantile levels due to data scarcity. Simulation experiments and an empirical example illustrate the performance and usefulness of the new model.
JOURNAL OF THE ROYAL STATISTICAL SOCIETY SERIES B-STATISTICAL METHODOLOGY
(2023)
Article
Engineering, Electrical & Electronic
Honglin Wen, Jie Gu, Jinghuan Ma, Lyuzerui Yuan, Zhijian Jin
Summary: A deep-learning forecasting model based on neural basis expansion analysis is proposed to improve short term load forecasting by narrowing the prediction interval. The model uses a doubly residual stacking strategy to decompose the forecasting task into sub-problems and applies conformal quantile regression for better theoretical coverage guarantee. Experiments show that the proposed model demonstrates improved performance in capturing load characteristics and providing narrow prediction intervals with nearly nominal coverage.
IEEE TRANSACTIONS ON SMART GRID
(2021)
Article
Mathematical & Computational Biology
Ryan Kelly, Kehui Chen
Summary: This paper applies conformal prediction techniques to construct prediction intervals in a multiple functional regression setting. A method with great modeling flexibility is developed using the Signature expansion. The resulting algorithm produces a closed form solution for a prediction set with guaranteed coverage.
STATISTICS AND ITS INTERFACE
(2022)
Article
Computer Science, Interdisciplinary Applications
Mohamed Noureldin, Tamer Abuhmed, Melike Saygi, Jinkoo Kim
Summary: A new probabilistic framework is proposed for distribution-free prediction interval (PI) of seismic responses in earthquake engineering. It overcomes the limitations of point prediction models and complexity of traditional probabilistic methods. The framework utilizes a few assumptions of probability distributions and requires no prior statistical distribution assumption for the PI.
COMPUTER-AIDED CIVIL AND INFRASTRUCTURE ENGINEERING
(2023)
Article
Engineering, Electrical & Electronic
Yan Zhou, Yonghui Sun, Sen Wang, Rabea Jamil Mahfoud, Hassan Haes Alhelou, Nikos Hatziargyriou, Pierluigi Siano
Summary: Accurate regional wind power prediction is crucial for the security and reliability of power systems. This study proposes a novel probabilistic prediction method based on composite conditional nonlinear quantile regression (CCNQR) to improve the performance of very short-term prediction intervals (PIs). The method considers static and dynamic differences as well as meteorological differences in wind power time series, and uses correlations as sample weights for prediction. The proposed method is verified to be superior to conventional methods through comparisons using real wind farm data.
JOURNAL OF MODERN POWER SYSTEMS AND CLEAN ENERGY
(2022)
Article
Statistics & Probability
Dhruv Medarametla, Emmanuel Candes
Summary: The study focuses on constructing confidence intervals for the conditional median without assuming anything about the distribution of the data. It proposes a method based on conformal prediction with theoretical guarantees, and establishes an equivalence between confidence intervals for the conditional median and response variable. The study also proves a lower bound independent of sample size for all distributions without point masses.
ELECTRONIC JOURNAL OF STATISTICS
(2021)
Article
Environmental Studies
Julian E. Lozano, Katarina Elofsson, Yves Surry
Summary: The study in Sweden found that wolves, lynxes, and brown bears may have negative impacts on hunting lease prices in the middle range, while no significant impacts were found in the lower quantiles. An additional wolf territory can result in a 21% reduction in hunting lease prices per hectare, while an additional lynx family group and brown bear individual can lead to reductions of 22.4% and 0.6% respectively.
Article
Computer Science, Artificial Intelligence
Nicolas Dewolf, Bernard De Baets, Willem Waegeman
Summary: In this independent comparative study, four classes of methods, including Bayesian methods, ensemble methods, direct interval estimation methods, and conformal prediction methods, are reviewed for their validity and calibration in the regression setting. Results on benchmark data sets show large fluctuations in performance across different domains. Conformal prediction can be used as a general calibration procedure for methods that deliver poor results without calibration.
ARTIFICIAL INTELLIGENCE REVIEW
(2023)
Article
Statistics & Probability
Sandra Siegfried, Lucas Kook, Torsten Hothorn
Summary: We introduce a generalized additive model for location, scale, and shape (GAMLSS) for distribution-free and parsimonious regression modelling. The model replaces strict parametric distribution with a transformation function and limits the number of linear or smooth model terms. Likelihood and score functions are derived for various types of observations. Various algorithms are used for model estimation and parameter interpretability is connected to model selection. A novel best subset selection procedure is proposed for simpler interpretation. Numerous applications are provided as examples and all analyses are reproducible using the tram add-on package to the R system.
AMERICAN STATISTICIAN
(2023)
Editorial Material
Statistics & Probability
Jelena Bradic, Yinchu Zhu
Editorial Material
Statistics & Probability
Jelena Bradic, Yinchu Zhu
Article
Economics
Simon C. Smith, Allan Timmermann, Yinchu Zhu
JOURNAL OF ECONOMETRICS
(2019)
Article
Economics
Ivana Komunjer, Yinchu Zhu
JOURNAL OF ECONOMETRICS
(2020)
Article
Statistics & Probability
Victor Chernozhukov, Kaspar Wuthrich, Yinchu Zhu
Summary: This study introduces new inference procedures for policy evaluation using counterfactual and synthetic control methods. By recasting the causal inference problem as a counterfactual prediction and structural breaks testing problem, and exploiting insights from conformal prediction and structural breaks testing, the study develops permutation inference procedures that accommodate modern high-dimensional estimators.
JOURNAL OF THE AMERICAN STATISTICAL ASSOCIATION
(2021)
Article
Economics
Yinchu Zhu, Allan Timmermann
Summary: This study establishes conditions for improving forecasting performance by rotating between a set of underlying forecasts and tracking their predictive accuracy using time-varying monitoring instruments. The properties of the monitoring instruments that are useful for identifying the best forecast at each point in time are characterized, reflecting the accuracy of the predictors and the strength of the monitoring instruments. Finite-sample bounds on forecasting performance are used to compute the expected loss of competing forecasts and guide the dynamic rotation strategy. Monte Carlo simulations and empirical applications demonstrate the potential gains from using conditioning information to rotate between forecasts.
JOURNAL OF ECONOMETRICS
(2022)
Article
Statistics & Probability
Jelena Bradic, Jianqing Fan, Yinchu Zhu
Summary: Understanding statistical inference under possibly nonsparse high-dimensional models has gained attention recently. The difficulty of the problem depends on the sparsity of the precision matrix rather than the regression coefficients. New concepts of uniform and essentially uniform nontestability allow studying the limitations of tests across a broad set of alternatives. These concepts lead to new minimax testability results that do not rely on the sparsity of the regression parameters. The study also reveals tradeoffs between testability and feature correlation.
ANNALS OF STATISTICS
(2022)
Editorial Material
Multidisciplinary Sciences
Victor Chernozhukov, Kaspar Wuethrich, Yinchu Zhu
PROCEEDINGS OF THE NATIONAL ACADEMY OF SCIENCES OF THE UNITED STATES OF AMERICA
(2023)
Article
Statistics & Probability
Victor Chernozhukov, Christian Hansen, Yuan Liao, Yinchu Zhu
Summary: This paper studies inference in linear models with a high-dimensional parameter matrix that can be well approximated by a spiked low-rank matrix. The framework covers a broad class of models that can accommodate matrix completion problems, factor models, varying coefficient models and heterogeneous treatment effects. The proposed procedure provides asymptotically normal inference and achieves the semiparametric efficiency bound.
ANNALS OF STATISTICS
(2023)