|
Abstract:
|
This article introduces an entirely novel nonparametric methodology for Generalized Linear Models (GLMs), through an effective transformation of the current framework. It is shown to be an extension of recent parametric advances giving results superior to it in various settings. Despite being nonparametric it does not necessarily need more iterations for convergence in comparison to the parametric version, if the underlying DGP is symmetric. If the underlying DGP is asymmetric it gives uniformly better prediction and inference performance to the existing methodologies compared. Furthermore, we present a new classification statistics utilizing which we show that it has better inference and classification performance than the parametric version, which is statistically significant especially if the DGP asymmetric. We further show that the methodology can outperform existing Artificial Intelligence and Machine Learning methods such as Neural Networks. In addition, we show that the methodology can be used to perform model diagnostics for any categorical model, a highly useful and novel result in the field. Finally we apply the method to various real world data, and discuss the findings.
|
