Performance Evaluation of Nonparametric Predictive Inference

非参数预测推理(NPI)是一种基于少量建模假设的频数统计框架，通过使用不确定性量化的不精确概率而成为可能。自20世纪90年代末以来，监管者发起并发展了NPI，并提出了广泛应用的理论和方法。由于基于NPI的方法不同于经典方法，如何评估它们的性能并不是一件轻而易举的事情，现有的评估方法不适合评估不精确的预测。在这个项目中，我们将进一步研究评估预测性能的机会，特别是考虑到不精确度的水平，只要整体预测结果良好，不精确度越低越好。这是一个为期4年的博士项目，第一个月将用来熟悉NPI的技术现状和一些应用领域。研究项目的核心将包括开发绩效评估方法，比较替代推理方法，并使用统计软件R。

Nonparametric Predictive Inference (NPI) is a frequentist statistics frameworkbased on only few modelling assumptions, made possible by the use ofimprecise probability for uncertainty quantification. The supervisors have initiatedand developed NPI since the late 90s, and theory and methods have beenpresented for a wide range of applications. As the NPI-based methods aredifferent from classical methods it is not trivial how we can assess their performance, existing methods to assess estimation performance are not suitableto assess imprecise predictions. In this project, we will investigate furtheropportunities to assess the predictive performance, in particular to take intoaccount the level of imprecision, with less imprecision being better as long as theoverall prediction results are fine. This is a 4-year PhD project, in which the firstmonths will be used to get familiar with the current state of the art of NPI andsome of its application areas. The core of the research project will consist ofdeveloping performance evaluation methods, comparison to alternative inferencemethods, and implementing these using the statistical software R.