Special Session 81: Improving climate and weather prediction through data-driven statistical modeling
Contents
We address the problem of predicting future measurements of a single time series from a set of multivariate predictor time series in an information theoretic framework. We investigate in how far this can be done optimally given the available information. Such an optimality criterion has to balance the effect of including too few or the wrong predictors (model-misspecification) and too many predictors (practical problem of overfitting) and develop a practical prediction algorithm to address these issues. The performance and challenges are demonstrated on multivariate nonlinear stochastic delay processes as well as on real climate data.