Stabilised weighted data subsampling for accelerated Bayesian inference in models with recursive likelihoods
Abstract
Bayesian inference for models with recursively defined likelihoods is computationally demanding, limiting scalability to large datasets. We propose a stabilised weighted subsampling methodology for accelerated Bayesian inference based on an unbiased estimator of the log-likelihood. By assigning higher sampling probabilities to early observations, the method reduces the effective depth of recursive likelihood evaluations and hence computational cost. This creates a challenging variance-cost trade-off: sampling probabilities that decay too slowly yield limited computational savings, while overly aggressive decay can substantially inflate estimator variance. We develop a stabilisation framework, supported by theory, that controls this trade-off and leads to a practical, theory-informed tuning strategy. We also derive an unbiased subsampling estimator of the log-likelihood gradient, enabling gradient-based Bayesian inference. The methodology can be embedded within a range of Bayesian computational algorithms. We illustrate its use in variational Bayes and subsampling Markov chain Monte Carlo for conditional volatility models, including leverage effects. Empirical results show substantial computational speed-ups relative to full-data methods while maintaining inferential accuracy. We also compare with recent stochastic gradient MCMC and divide-and-conquer MCMC methods for temporally dependent data, observing favourable empirical performance.
Related
- Vine Copula VAR:From Recursive Margins to Joint Forecast Inference
- Recursive-Head Geometry and Order-Free Efficient Inference in Finite-State Nested Markov Models
- Inference for Newton Methods with Accelerated Sketch-and-Project via Random Scaling
- Central limit theorems for high-dimensional ridge regression under uniform subsampling
- Optimal subsampling approach for Cox models under large-scale length-biased right-censored data
BibTeX
@misc{quiroz2026stabilised,
title = {{Stabilised weighted data subsampling for accelerated Bayesian inference in models with recursive likelihoods}},
author = {Matias Quiroz and Aishwarya Bhaskaran and Zixuan Wang and Thomas Goodwin},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2605.13397},
eprint = {2605.13397},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2605.13397},
}
RIS
TY - UNPB
TI - Stabilised weighted data subsampling for accelerated Bayesian inference in models with recursive likelihoods
AU - Matias Quiroz
AU - Aishwarya Bhaskaran
AU - Zixuan Wang
AU - Thomas Goodwin
PY - 2026
DA - 2026/10/10
PB - arXiv
DO - 10.48550/arxiv.2605.13397
UR - https://arxiv.org/abs/2605.13397
AB - Bayesian inference for models with recursively defined likelihoods is computationally demanding, limiting scalability to large datasets. We propose a stabilised weighted subsampling methodology for accelerated Bayesian inference based on an unbiased estimator of the log-likelihood. By assigning higher sampling probabilities to early observations, the method reduces the effective depth of recursive likelihood evaluations and hence computational cost. This creates a challenging variance-cost trade-off: sampling probabilities that decay too slowly yield limited computational savings, while overly aggressive decay can substantially inflate estimator variance. We develop a stabilisation framework, supported by theory, that controls this trade-off and leads to a practical, theory-informed tuning strategy. We also derive an unbiased subsampling estimator of the log-likelihood gradient, enabling gradient-based Bayesian inference. The methodology can be embedded within a range of Bayesian computational algorithms. We illustrate its use in variational Bayes and subsampling Markov chain Monte Carlo for conditional volatility models, including leverage effects. Empirical results show substantial computational speed-ups relative to full-data methods while maintaining inferential accuracy. We also compare with recent stochastic gradient MCMC and divide-and-conquer MCMC methods for temporally dependent data, observing favourable empirical performance.
ER -