Forecasting Recovery Rates on Non-Performing Loans with Machine Learning

Bellotti, Tony; Brigo, Damiano; Gambetti, Paolo; Vrins, Frédéric

doi:10.2139/ssrn.3434412

Cited by 10 publications

(4 citation statements)

References 28 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Several empirical studies were interested in the method of management of NPLs. For example, Bellotti, Brigo, Gambetti, and Vrins (2019) have performed a set of regression techniques and machine learning algorithms to predict the recovery rates on NPLs. The sample used in this study is based on a large transaction of NPLs between European banks.…”

Section: Related Literaturementioning

confidence: 99%

Is there a threshold effect in the liquidity risk–non‐performing loans relationship? APSTRapproach forMENAbanks

2020

View full text Add to dashboard Cite

This study analyzes the nonlinear relationship between liquidity risk and nonperforming loans (NPLs) for a sample of MENA banks over the period 2004-2017. Results of the Panel Smooth Transition Regression model indicate that there is a threshold effect in the liquidity risk and NPLs relationship. More specifically, we found that above the threshold of 73.10% for loans to deposits ratio, liquidity risk significantly increases the level of NPLs. However, below the threshold of 87.61% for liquid assets to deposits and short-term funding ratio, the NPLs are also significantly and positively correlated with the liquidity risk. Furthermore, we found that NPLs are more sensitive to bank performance, bank capital, bank size, international financial crisis, and the inflation rate. However, no significant effect was found for the impact of ownership concentration and board characteristics either below or above the thresholds.

show abstract

Section: Related Literaturementioning

confidence: 99%

Is there a threshold effect in the liquidity risk–non‐performing loans relationship? APSTRapproach forMENAbanks

2020

View full text Add to dashboard Cite

show abstract

“…Machine learning (ML) has recently emerged as a critical predictive mechanism in a number of fields, including in education [19], [20], fraud detection [21] and medical [22]. Also, a number of studies [23]- [25] have demonstrated that ML can produce findings that are more accurate when predicting NPLs. For instance, Bellotti et al [23] uses a private database from a European debt collection firm to compare the results of a diverse range of regression approaches and ML algorithms for forecasting recovery rates on nonperforming loans.…”

Section: Introductionmentioning

confidence: 99%

“…Also, a number of studies [23]- [25] have demonstrated that ML can produce findings that are more accurate when predicting NPLs. For instance, Bellotti et al [23] uses a private database from a European debt collection firm to compare the results of a diverse range of regression approaches and ML algorithms for forecasting recovery rates on nonperforming loans. The results show that rule-based algorithms such as cubist, boosted trees and random forests perform significantly better than other approaches.…”

Section: Introductionmentioning

confidence: 99%

The nexus of corruption and non-performing loan: machine learning approach

Masrom,

Rahman,

Shukri

et al. 2023

IJEECS

View full text Add to dashboard Cite

<span>Banking institutions around the world are facing a serious problem with non-performing loans (NPLs), which can jeopardize their financial stability and hinder their ability to issue new loans. The issue of NPLs has been linked to corruption, which has emerged as one of the contributing factors. Given the scarcity of research on the use of machine learning (ML) techniques to examine the relationship between corruption and NPLs, this paper provides an empirical evaluation of various ML algorithms for predicting NPLs. Besides ML performance comparisons, this paper presents the analysis of ML features importance to justify the effect of corruption factor in the different ML algorithms for predicting NPLs. The results indicated that most of the tested ML algorithms present good ability in the prediction models at accuracy percentages above 70% but corruption index has contributed very minimal effect to the ML performances. The most outperformed ML algorithm in the different proposed settings is random forest. The framework of this research is highly reproducible to be extended with a more in-depth analysis, particularly on problems of NPL as well as on the ML algorithms.</span>

show abstract

“…In particular, the work confirms, with (Khieu, et al, 2012), that loan's characteristics are more significant determinants of the recovery rate than are borrower characteristics prior to default. On the other hand, we were not able to inquire if predictors referring to the recovery process of the processor, or the bank prior to portfolio sale, were relevant as suggested by (Bellotti, et al, 2019), since such information was not available. The approach adopted mirrors the standard Loss Given Default (LGD) framework which splits the default event from loss estimates and was designed independently of existing literature.…”

Section: Introductionmentioning

confidence: 99%