Selected work

Representative Papers

Learning Bidirectional Causal Interactions with Heteroscedastic Neural Networks

Jul 24, 2026

Traditional methods struggle to identify bidirectional causal relationships between endogenous variables from observational data. This work proposes SEM-DNN, a novel approach that integrates conditional heteroskedasticity structures with deep neural networks to jointly model nonlinear structural equations and variance dependencies in disturbance terms. By enforcing diagonalization of the conditional covariance matrix, the method achieves identification of bidirectional causality without requiring instrumental variables, thereby guaranteeing unique identifiability of structural parameters. SEM-DNN substantially outperforms existing parametric, kernel-based, and decoupled neural network methods under conditions of nonlinearity, high-dimensional confounding, and non-Gaussian disturbances. Empirical application to price–sales feedback analysis in breakfast cereals demonstrates both its causal identification capability and the effectiveness of residual diagonalization.

0 citationsRead paper

A Theory of Bootstrap Coverage Calibration for Generalized Posterior Credible Sets

Jun 24, 2026

Generalized posterior credible sets typically lack asymptotic frequentist coverage guarantees. This work proposes calibrating their coverage by estimating coverage probabilities via the bootstrap and selecting a scalar learning rate to achieve the desired nominal level. Leveraging Edgeworth expansions, we show that coverage error arises from two sources: corrections to the sampling distribution of the estimator and posterior-induced adjustments to the center, boundary, and shape of the credible set. We further establish that a scalar learning rate can yield globally valid calibration only when the posterior and sampling covariance matrices are proportional. Under fixed-dimensional asymptotics, regularity conditions, and local identifiability, we prove—using Edgeworth expansions, classical asymptotic theory, and stochastic approximation—that the solution to the bootstrap coverage equation is consistent, revealing that the procedure essentially implements a scale correction tailored to a specific confidence level.

0 citationsRead paper

Quasi-Bayesian Local Projection Instrumental-Variables Method: Application to Renewable Energy and Electricity Prices

May 15, 2026

This study addresses the instability of conventional local projection instrumental variables (LP-IV) estimators in finite samples, particularly the large estimation errors in medium- to long-run impulse responses and the difficulty of conducting joint inference. The authors propose a quasi-Bayesian LP-IV estimator that constructs a moment-based quasi-posterior distribution from the generalized method of moments (GMM) objective function and, for the first time, incorporates a roughness-penalty prior to impose smoothness constraints on multi-horizon responses. While preserving the standard first-order asymptotic properties of traditional LP-IV, the proposed method substantially improves finite-sample stability and enables joint inference via simultaneous confidence bands. Monte Carlo simulations demonstrate markedly lower root mean squared errors, especially at longer horizons, and an empirical application to the Danish electricity market further corroborates its practical relevance.

0 citationsRead paper

Inferring Microscopic Explanatory Structures from Observational Constraints via Large Deviations

Feb 06, 2026

This study addresses the challenge of inferring plausible microscopic explanations solely from macroscopic observational constraints, without relying on prior assumptions about microstructure or dynamics. The problem is formulated as a constrained large deviations optimization: among all microscopic configurations compatible with the given macroscopic constraints, the statistically most typical solution is selected by minimizing the relative entropy with respect to a symmetric reference measure uniquely determined by the measurement setup. Remarkably, this approach requires no predefined notion of order yet spontaneously yields structured interpretations—such as ordered relationships—as emergent properties. The work demonstrates that even under a fully permutation-symmetric reference measure, ordered microscopic structures can naturally arise as typicality-optimal solutions.

0 citationsRead paper

Modified Delayed Acceptance MCMC for Quasi-Bayesian Inference with Linear Moment Conditions

Nov 21, 2025

This paper addresses quasi-Bayesian inference under high-dimensional linear moment conditions. We propose a computationally efficient and numerically stable delayed-acceptance MCMC framework. Methodologically, we (1) construct a structured surrogate target kernel and an approximate conditional posterior proposal, explicitly exploiting the linearity of the moment conditions; (2) develop two complementary implementations—Exact, which maximizes single-step sampling efficiency, and Approx, which avoids high-dimensional matrix inversion to improve throughput and numerical stability; and (3) extend the framework to a class of risk-based Bayesian models grounded in first-order moment conditions. Empirical evaluations on heteroskedastic regression and instrumental variable applications demonstrate substantial improvements over standard MCMC: the Approx variant achieves optimal throughput in large-scale settings, while the Exact variant delivers significantly higher effective sample size per unit computation time and markedly improved parameter estimation accuracy.

0 citationsRead paper

Recent publications

Latest Papers

Learning Bidirectional Causal Interactions with Heteroscedastic Neural Networks

Jul 24, 2026

Traditional methods struggle to identify bidirectional causal relationships between endogenous variables from observational data. This work proposes SEM-DNN, a novel approach that integrates conditional heteroskedasticity structures with deep neural networks to jointly model nonlinear structural equations and variance dependencies in disturbance terms. By enforcing diagonalization of the conditional covariance matrix, the method achieves identification of bidirectional causality without requiring instrumental variables, thereby guaranteeing unique identifiability of structural parameters. SEM-DNN substantially outperforms existing parametric, kernel-based, and decoupled neural network methods under conditions of nonlinearity, high-dimensional confounding, and non-Gaussian disturbances. Empirical application to price–sales feedback analysis in breakfast cereals demonstrates both its causal identification capability and the effectiveness of residual diagonalization.

0 citationsRead paper

A Theory of Bootstrap Coverage Calibration for Generalized Posterior Credible Sets

Jun 24, 2026

Generalized posterior credible sets typically lack asymptotic frequentist coverage guarantees. This work proposes calibrating their coverage by estimating coverage probabilities via the bootstrap and selecting a scalar learning rate to achieve the desired nominal level. Leveraging Edgeworth expansions, we show that coverage error arises from two sources: corrections to the sampling distribution of the estimator and posterior-induced adjustments to the center, boundary, and shape of the credible set. We further establish that a scalar learning rate can yield globally valid calibration only when the posterior and sampling covariance matrices are proportional. Under fixed-dimensional asymptotics, regularity conditions, and local identifiability, we prove—using Edgeworth expansions, classical asymptotic theory, and stochastic approximation—that the solution to the bootstrap coverage equation is consistent, revealing that the procedure essentially implements a scale correction tailored to a specific confidence level.

0 citationsRead paper

Quasi-Bayesian Local Projection Instrumental-Variables Method: Application to Renewable Energy and Electricity Prices

May 15, 2026

This study addresses the instability of conventional local projection instrumental variables (LP-IV) estimators in finite samples, particularly the large estimation errors in medium- to long-run impulse responses and the difficulty of conducting joint inference. The authors propose a quasi-Bayesian LP-IV estimator that constructs a moment-based quasi-posterior distribution from the generalized method of moments (GMM) objective function and, for the first time, incorporates a roughness-penalty prior to impose smoothness constraints on multi-horizon responses. While preserving the standard first-order asymptotic properties of traditional LP-IV, the proposed method substantially improves finite-sample stability and enables joint inference via simultaneous confidence bands. Monte Carlo simulations demonstrate markedly lower root mean squared errors, especially at longer horizons, and an empirical application to the Danish electricity market further corroborates its practical relevance.

0 citationsRead paper

Inferring Microscopic Explanatory Structures from Observational Constraints via Large Deviations

Feb 06, 2026

This study addresses the challenge of inferring plausible microscopic explanations solely from macroscopic observational constraints, without relying on prior assumptions about microstructure or dynamics. The problem is formulated as a constrained large deviations optimization: among all microscopic configurations compatible with the given macroscopic constraints, the statistically most typical solution is selected by minimizing the relative entropy with respect to a symmetric reference measure uniquely determined by the measurement setup. Remarkably, this approach requires no predefined notion of order yet spontaneously yields structured interpretations—such as ordered relationships—as emergent properties. The work demonstrates that even under a fully permutation-symmetric reference measure, ordered microscopic structures can naturally arise as typicality-optimal solutions.

0 citationsRead paper

Modified Delayed Acceptance MCMC for Quasi-Bayesian Inference with Linear Moment Conditions

Nov 21, 2025

This paper addresses quasi-Bayesian inference under high-dimensional linear moment conditions. We propose a computationally efficient and numerically stable delayed-acceptance MCMC framework. Methodologically, we (1) construct a structured surrogate target kernel and an approximate conditional posterior proposal, explicitly exploiting the linearity of the moment conditions; (2) develop two complementary implementations—Exact, which maximizes single-step sampling efficiency, and Approx, which avoids high-dimensional matrix inversion to improve throughput and numerical stability; and (3) extend the framework to a class of risk-based Bayesian models grounded in first-order moment conditions. Empirical evaluations on heteroskedastic regression and instrumental variable applications demonstrate substantial improvements over standard MCMC: the Approx variant achieves optimal throughput in large-scale settings, while the Exact variant delivers significantly higher effective sample size per unit computation time and markedly improved parameter estimation accuracy.

0 citationsRead paper