题 目：Prediction under abrupt changes or smooth evolutions
主讲人：Jie Ding, Assistant Professor, Department of statistics, university of Minnesota
In this talk I will introduce a new method referred to as kinetic prediction for predicting sequential data with unknown abrupt changes in their data generating distributions. Based on Kolmogorov-Tikhomirov’s epsilon entropy, we propose a concept called epsilon-predictability that quantifies the size of a model class and the maximal number of structural changes that guarantee the achievability of asymptotically optimal prediction. Moreover, for parametric distribution families, the aforementioned kinetic prediction with discretized function spaces is extended to its counterpart with continuous function spaces, which naturally leads to an efficient sequential Monte Carlo implementation. The method is also extended to handle smooth evolutions of the underlying data generating distribution. The applicability of the introduced methods will be illustrated by both synthetic and real-world data.