Algorithmic Decisions in Debt Collection

How to cite

Zhou, Yijun. 2024. "Algorithmic Decisions in Debt Collection." AEA RCT Registry. AEARCTR-0013905. https://doi.org/10.1257/rct.13905-1.1

Download: BibTeX · RIS — for reference managers without a browser translator. Zotero and most others pick up this page's citation metadata directly.

Yijun Zhou*

* Primary investigator

Registry ID
AEARCTR-0013905
DOI
https://doi.org/10.1257/rct.13905-1.1
First registered
Last updated
Status
Completed
Trial dates
2020-06-01 – 2021-01-01
Countries
Netherlands ()
JEL codes
D14, G02, G23, G51, O33
Registry record
http://www.socialscienceregistry.org/trials/13905

Abstract

This paper examines the role of artificial intelligence (AI) in facilitating the non-judicial collection process of delinquent consumer debt. Leveraging a randomized field experiment in the Netherlands, we show that algorithmic calling decisions achieve higher repayment rates with fewer collection calls compared with human collection officers. Uncovering the black box of AI, we find that it extracts predictive signals from unstructured notes compiled by collectors. These signals not only predict whether the delinquent borrowers would repay during the non-judicial collection process, but also shed light on the underlying motivations or impediments of delinquent borrowers' repayment behavior.

Intervention

The experiment was independently designed and implemented by our data provider (i.e., the debt collection agency) to examine the effectiveness of algorithmic calling decisions. In June 2020, The debt collection agency randomly assigned these 7,839 borrowers into two groups. The first group, consisting of 3,885 borrowers (the control group), receives calls determined by human collection officers. The second group, comprising of 3,954 borrowers (the treated group), receives calls based algorithmic decisions during the whole collection process. Contacting/calling in both groups is made by the same team of human calling agents. In this randomized experiment, the "treatment" is defined not by whether borrowers receive calls from calling agents, but by who decide those calls. By the end of the collection process, data was collected for these borrowers' repayment behavior.

Experimental design

In June 2020, The debt collection agency randomly assigned these 7,839 borrowers into two groups. The first group, consisting of 3,885 borrowers (the control group), receives calls determined by human collection officers. The second group, comprising of 3,954 borrowers (the treated group), receives calls based algorithmic decisions during the whole collection process. Contacting/calling in both groups is made by the same team of human calling agents. In this randomized experiment, the "treatment" is defined not by whether borrowers receive calls from calling agents, but by who decide those calls. By the end of the collection process, data was collected for these borrowers' repayment behavior.

Design details

Randomization method
Randomization done electronically by the debt collection agency.
Randomization unit
Randomization was at the borrower level.
Clusters
7,839 borrowers
Observations
7,839 borrowers
Arms
3,954 borrowers
Primary outcomes
Repayment behavior of borrowers

Cited by

No citation data available for this registration yet.

Keywords