Generalised Bayesian model selection

Speaker: Dr Ritabrata Dutta, University of Warwick

Abstract: Scritly proper scoring rules (SPSR), e.g. Energy score, Kernel score, are widely used as diagnostics of probabilistic forecasting. As we can derive statistical divergences from SPSRs, recently they have become popular for the purpose of inference in both frequentist (as minimum scoring rules estimator) and Bayesian framework (as generalised posterior) in the absence of tractable likelihood functions via replacement of negative log-likelihoods with SPSRs. Here we explore a generalised Bayesian model selection framework for models without tractable likelihood functions, using Bayes factor derived from generalised posteriors. To compute the generalised marginal evidence under each model, we propose a path sampling based estimator in conjunction with sequential Monte Carlo sampling scheme. We study the consistency of the proposed procedure theoretically and using simulation studies, finally illustrating their use for the choice of challenging biological models without tractable likelihood functions.  

Biography: Ritabrata ‘Rito’ Dutta, Reader in the Department of Statistics (Warwick), works on robust simulation-based inference of generative models for probabilistic prediction. His collaboration with ECMWF on projects funded by ECMWF and the Turing Institute focuses on developing robust diagnostic tools to evaluate weather forecast models and provides foundation for training data-driven ML models for weather forecasting and downscaling via diagnostics tools like CRPS, which has become the go-to tool for all existing models. He is one of the pioneers for developing AI models for weather prediction downscaling tasks. He has (co-)led funded research totalling over £1.5M (EPSRC EP/V025899/1, EPSRC EP/T017112/1, NERC NE/T00973X/1), covering optimal lock-down strategies during COVID-19, predicting fish stock movements in the English Channel, or quantifying uncertainties in long-form text generation. 

Host: Prof. Deborah Sulem