B1265
Title: Towards black-box parameter estimation
Authors: Amanda Lenzi - University of Edinburgh (United Kingdom) [presenting]
Abstract: Statistical inference is at the core of modelling, prediction, and simulation. However, models designed to express complex dependence mechanisms and large volumes of data are computationally challenging using classical inference techniques, limiting their usefulness. For example, this applies to finance or climate science datasets, where skewness and jumps are commonly present, and likelihood computation is impossible even with small datasets due to unknown normalizing constants. Deep learning-based procedures are presented that estimate parameters of statistical models for which simulation is easy, but likelihood computation is challenging. Due to their amortized nature, these estimators are fast, likelihood-free, and amenable to fast bootstrap-based uncertainty quantification. The applicability of the proposed approaches is demonstrated to quickly and optimally obtain estimates and confidence intervals for parameters from non-Gaussian models with complex spatial and temporal dependencies.
