Higher-Order Marginal Likelihood Approximation for Nonlinear State Space Models

Joshua C. C. Chan; Wei Zhang

Home/Archive/arXiv

arXiv · 2610.34629

Higher-Order Marginal Likelihood Approximation for Nonlinear State Space Models

Joshua C. C. Chan·Wei Zhang

View original source ↗Read source document ↗Export metadata ↓

Abstract

Bayesian comparison of nonlinear state space models requires integrating over a growing latent state path and a low-dimensional parameter vector. We propose a marginal likelihood approximation that integrates out the states with a numerical or simulation-based evaluator and applies a second-order corrected Laplace approximation to the integral over the static parameters. The correction is expressed through smoothing cumulants of the complete-data log density and, under derivative-growth and posterior-concentration conditions, reduces the log marginal likelihood error from $O(T^{-1})$ to $O(T^{-2})$. In stochastic volatility experiments, it removes 64\% to 99\% of the Laplace error at far lower cost than simulation-based estimators. Applications to bivariate exchange-rate volatility models and to a sample of $T = 19{,}266$ five-minute stock returns find the strongest support for heavy-tailed observations and for a transitory volatility factor.

Explore related subjects

econ.EM

Keep this discovery

Explore connections, maps & timelines

BibTeX RIS

Joshua C. C. Chan, Wei Zhang. 2026-10-25. Higher-Order Marginal Likelihood Approximation for Nonlinear State Space Models. https://arxiv.org/abs/2610.34629

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

KEEP EXPLORING

Related papers

arXiv2026

Valid Post-Contextual Bandit Inference

Contextual multi-armed bandits (CMAB) are increasingly used as adaptively sampled experiments, after which researchers test hypotheses about treatment effects. We derive the Hájek-Le Cam limit experiment for multi-armed contextual bandits under equal-arms (weak-signal) asymptotics. It is Locally Asymptotically Quadratic, but not a (mixed) Gaussian shift experiment. Girsanov's theorem gives it a structural representation in stochastic differential equations. Together with a general weak convergence result, this yields the asymptotic null distribution and local asymptotic power of often-used test statistics. Two kinds of nuisance parameters threaten size control: the common reward level of the arms of interest and the rewards of the arms not of interest. The first threat is removed by translation-invariant sampling. We propose Regularized Translation-Invariant (RTI) versions of Thompson, tempered-greedy, tempered-(Lin)UCB, and exploration sampling. The second threat calls for appropriate pivots. The structural representation of the limit experiment is key to overcome the inferential challenges arising from adaptive sampling. Simulations corroborate the theory.

econ.EM↗

arXiv2026

Evaluating A/B Testing Methodologies via Sample Splitting: Theory and Practice

We develop a theoretical framework for sample splitting in A/B testing environments, where data for each test are partitioned into two splits to measure methodological performance when the true impacts of tests are unobserved. We show that sample-split estimators are generally biased for full-sample performance but consistently estimate sample-split analogues of it. We derive their asymptotic distributions, construct valid confidence intervals, and characterize the bias-variance trade-offs underlying sample-split design choices. We validate our theoretical results through simulations and provide implementation guidance for A/B testing products seeking to evaluate new estimators and decision rules.

econ.EM↗

arXiv2026

Bias Correction and Inference for Panel Quantile Regression

We develop a large-$n$, large-$T$ theory for the unsmoothed fixed-effects quantile regression estimator, paralleling Hahn and Newey (2004) for smooth nonlinear panel models. An exact rank-score identity for the profiled criterion, with a convexity argument, delivers the leading incidental parameter bias of the slope estimator, of order $T^{-1}$, in closed form. When $n$ and $T$ grow at the same rate, the estimator is asymptotically normal but in general not centred, and conventional intervals are invalid. Admissible growth of $n$ relative to $T$, up to $n=o(T^3)$, depends on the regularity of the conditional error densities. A split-panel jackknife removes the leading bias throughout this range and yields valid inference. In serially dependent panels whose dependence decays geometrically or polynomially, the bias acquires a term from the dependence between regressors and errors of different periods, the jackknife remains valid with contiguous but not with interleaved half-panels, and clustered standard errors are valid.

econ.EM↗