A0206
Title: Beyond the sample: Bootstrap ex ante accuracy estimation under any regression model
Authors: Alicja Wolny-Dominiak - University of Economics in Katowice (Poland) [presenting]
Tomasz Zadlo - University of Economics in Katowice (Poland)
Abstract: Bootstrap methods are widely used in data analysis for various purposes, including the estimation of sampling distributions, assessing the precision of the estimation, estimating and correcting bias, and improving hypothesis testing. They are also popular in time series and spatial prediction, utilizing both cross-sectional and longitudinal data, to estimate the ex-ante prediction accuracy. However, the model-based bootstrap algorithms currently available are designed for specific parametric models. We propose a new bootstrap algorithm that can be applied to any model, enabling the generation of both sample and out-of-sample data. This allows for the estimation of ex-ante prediction accuracy for forecasts derived from both parametric and nonparametric models. Furthermore, our approach facilitates the estimation of ex-ante prediction accuracy under machine learning models, where traditional methods, such as k-fold cross-validation, assess the ex-post prediction accuracy. Our proposal is extensively tested in two Monte Carlo simulation studies. In the first study, we analyse its performance under a linear mixed model, comparing it with traditional bootstrap methods, such as parametric and residual bootstrapping, which have been shown to possess good properties in this case. The second simulation study examines the properties of our method under a nonparametric model, where estimating ex-ante prediction accuracy presents a significant challenge.
