Stabilised weighted data subsampling for accelerated inference in models with recursive likelihoods

Abstract

Inference for models with recursively defined likelihoods is computationally demanding, limiting scalability to large datasets. This talk introduces a stabilised weighted data subsampling approach that accelerates likelihood evaluation via an unbiased estimator of the log-likelihood based on non-uniform sampling probabilities. The method assigns higher sampling probabilities to early observations to reduce the effective depth of recursive computations. We propose a stabilisation framework that addresses the variance and computational pathologies that arise when the sampling probabilities decay either too slowly or too aggressively, enabling principled tuning of the method's hyperparameters. The resulting estimators can be embedded within data subsampling-based inference algorithms. Applications to conditional volatility models demonstrate substantial computational gains relative to both full-data methods and uniform subsampling.

This is joint work with Aishwarya Bhaskaran (University of New South Wales), Zixuan Wang (University of Technology Sydney), and Thomas Goodwin (University of Technology Sydney).