Sampling by Reversing The Landmarking Process

TL;DR

This paper introduces a novel backward landmark sampling method that, with a flexible weighting scheme, outperforms traditional forward landmark sampling in reducing sampling errors, demonstrated through mortgage data analysis.

Contribution

The paper proposes a backward landmark sampling technique with a progressive weighting scheme, offering an alternative to forward landmarking methods.

Findings

Backward landmark sampling has smaller errors than forward methods.

The method is effective on real mortgage data.

Flexible weighting improves sampling accuracy.

Abstract

Variations of the commonly applied landmark sampling are presented. These samplings are "forward" landmarking such that each stack of data is created by first selecting landmarks and then including the subsequent observations of the selected landmarks. Unlike these forward landmarking samplings, a "backward" or "reverse" landmarking is proposed with a flexible "progressive" weighting on selecting different types of events and non-events. The backward landmark sample is compared with those forward landmark samples with a real world mortgage data. Results show that the backward landmark sample has smaller sampling errors than of those forward landmark samples.