This talk has captions. You can remove these by pressing CC on the video toolbar.
Name: Jonas Brehmer
Talk Title: Using scoring functions to evaluate point process forecasts
Abstract: Many decision makers in industry or public institutions rely on forecasts of certain random quantities when choosing among alternative actions and assessing the associated risks. If relevant events cannot be measured at fixed points, but occur randomly in space and/or time, point processes arise as natural models in many applications, e.g. epidemiology, seismology, or quantitative criminology. Such models are often used to create point process-based forecasts, for instance (conditional) intensities, or higher order summary statistics such as pair correlations. Usually, decision makers will face a number of different predictions concerning these quantities, making a comparison of their accuracy crucial for well-founded decisions.
A principled approach to comparative forecast evaluation relies on scoring or loss functions, which assign a real number to each pair of forecast and realized observation of a random variable. We establish some general results which transfer the idea of scoring functions to the point process setting. This leads to a novel comparative evaluation method for point process-based forecasts and provides a new perspective on several existing ones. Since our approach focuses on relative performance, it can be used as a model selection technique which complements existing goodness-of-fit tests.
This talk is an invited talk at EVA 2021. View the programme here.