A0827
Title: Nonlinear autoregressive models for functional time series with Bayesian additive regression trees
Authors: Eoghan O Neill - Erasmus University Rotterdam (Netherlands) [presenting]
Maria Grith - Erasmus University Rotterdam (Netherlands)
Anastasija Tetereva - Erasmus University Rotterdam (Netherlands)
Jiahao Cao - The University of Texas Health Science Center at Houston (United States)
Guanyu Hu - The University of Texas Health Science Center at Houston (United States)
Abstract: The purpose is to introduce Bayesian additive regression tree models for function-on-function regression. The outcome function is modelled as a linear combination of data-adaptive basis functions. The coefficients of basis functions are determined by sums of trees that can split on both scalar and functional variables, including the lag of the dependent variable. Splitting rules for functions are defined by inner products between functional covariates and linear combinations of fixed basis functions, distinct from the aforementioned data-adaptive basis. A prior is considered on functional splits, allowing Markov chain Monte Carlo tree samples to adapt to the data by placing higher probability on selecting a subset of relevant basis functions for a splitting rule. The forecasting performance of the method is evaluated in an application to option pricing implied volatility surface data.
