Optimal Predictor Selection for High-dimensional Nonparametric Forecasting

This research investigates nonparametric additive models as a statistical tool for forecasting time series data. It introduces the Sparse Multiple Index (SMI) Modelling algorithm, which automatically identifies relevant information and discards less important information in complex real-world problems, without requiring expert knowledge. The algorithm takes historical information as input, groups that information through algorithmic reasoning, and captures potentially complex relationships between past and future values, resulting in an optimised model with improved predictive accuracy. The thesis also presents a new method, Conformal Bootstrap, for quantifying forecast uncertainty, and provides open-source software to support broader application of the proposed methods.