A class of models for Bayesian predictive inference (Q2214266)

The authors are studying the Bayesian framework enabling to make sequential prediction an a sequence \(X_1,X_2,\dots\) of random observations. Main point is to assign a predictive distribution \(\sigma_n(\cdot)=P(X_{n+1}\in\cdot \mid X_1,\dots X_n)\) directly without passing through the classical prior/posterior Bayesian scheme. To that purpose a class of predictive distributions \(\Sigma\) is introduced and its properties studied. Concerning the data sequence \(\{X_n\}\), it is assumed to be exchangeable. Both standard (i.e. based on the prior \(\pi\) and calculated posterior \(\pi_n\)) and non-standard procedures enabling an appropriate strategy for selecting~\(\sigma\) are assumed. The authors concentrate on the NSA approach, where the inferrer selects his/her predictive distribution~\(\sigma_n\). The Ionescu-Tulcea theorem is one of the key's tools. In addition to introducing~\(\Sigma\), main contributions are Theorems 4, 5, 18 and~20. The former state that \(\{X_n\}\) is c.i.d. under \(P_{\sigma}\ \forall \sigma\in\Sigma\), while the latter deal with the asymptotics of \(\sigma_n\). Many interesting examples illustrate the approach. They support the fact that~\(\Sigma\) is rich enough to cover a wide range of real problems.