Heteroskedastic Gaussian processes for ecological forecasting

Loading...
Thumbnail Image

TR Number

Date

2026-10-09

Journal Title

Journal ISSN

Volume Title

Publisher

Virginia Tech

Abstract

Modern stochastic simulators and ecological time series share a common challenge: they exhibit input dependent noise. In such situations, accurate predictions into the future along with principled uncertainty are essential. Heteroskedastic Gaussian processes (hetGPs) are a flexible and nonparametric modeling technique wherein, two processes are inferred simul- taneously – the mean response and the latent noise levels corresponding to input locations. In this dissertation, I show how hetGPs are versatile and can be used in both small and large scale regimes. First, I show how hetGPs outperform other models in data-driven forecasting of sparse and irregularly sampled tick densities across several different locations. Then, I develop a Bayesian and scalable hetGP (bhetGP) to handle the scale of modern stochastic simulation campaigns without undercutting UQ. Next, I use this novel bhetGP for cali- bration of a computer model designed to forecast lake temperatures. Finally, I provide an extension to hetGPs where the noise can only be modeled for a subset of dimensions re- ducing computational cost and complexity. An open-source implementation for the bhetGP method and the extension is provided as the bhetGP R package on CRAN. Together these contributions establish hetGPs as a principled, scalable, and broadly applicable framework for ecological forecasting – one where uncertainty is a core deliverable.

Description

Keywords

uncertainty quantification, Bayesian inference, Vecchia, replication, input dependent noise, sparse time series, stochastic simulations

Citation