Prediction of a complex system with few data: Evaluation of the effect of model structure and amount of data with dynamic bayesian network models

Abstract

A major challenge in environmental modeling is to identify structural changes in the ecosystem across time, i.e., changes in the underlying process that generates the data. In this paper, we analyze the Baltic Sea food web in order to 1) examine potential unobserved processes that could affect the ecosystem and 2) make predictions on some variables of interest. To do so, dynamic Bayesian networks with different setups of hidden variables (HVs) were built and validated applying two techniques: rolling-origin and rolling-window. Moreover, two statistical inference approaches were compared at regime shift detection: fully Bayesian and Maximum Likelihood Estimation. Our results confirm that, from the predictive accuracy point of view, more data help to improve the predictions whereas the different setups of HVs did not make a critical difference in the predictions. Finally, the different HVs picked up patterns in the data, which revealed changes in different parts of the ecosystem.