Credit Risk Estimation Research Articles

Comparison of the impact of dimensionality reduction and data splitting on classification performance in credit risk assessment

Credit risk assessment (CRA) plays an important role in credit decision-making process of financial institutions. Today, developing big data analysis and machine learning methods have marked a new era in credit risk estimation. In recent years, using machine learning methods in credit risk estimation has emerged as an alternative method for financial institutions. The past demographic and financial data of the person whose CRA will be performed is important for creating an automatic artificial intelligence credit score prediction model based on machine learning. It is also necessary to use features correctly to create accurate machine learning models. This article aims to investigate the effects of dimensionality reduction and data splitting steps on the performance of classification algorithms widely used in the literature. In our study, dimensionality reduction was performed using Principal Component Analysis (PCA). Random Forest (RF), Logistic Regression (LR), Decision Tree (DT), Naive Bayes (NB) algorithms were chosen for classification. Percentage splitting (PER, 66–34%) and k-fold (k = 10) cross-validation techniques were used when dividing the data set into training and test data. The results obtained were evaluated with accuracy, recall, F1 score, precision, and AUC metrics. German data set was used in this study. The effect of data splitting and dimension reduction on the classification of CRA systems was examined. The highest ACC in PER and CV data splitting was obtained with the RF algorithm. Using data splitting methods and PCA, the highest accuracy was observed with RF and the highest AUC with NB, with 13 PCs in which 80% of the variance was obtained. As a result, the data set consisting of a total of 20 features, expressed by 13 PCs, achieved similar or higher success than the results obtained from the original data set.

Discrete-Time Survival Models with Neural Networks for Age–Period–Cohort Analysis of Credit Risk

Survival models have become popular for credit risk estimation. Most current credit risk survival models use an underlying linear model. This is beneficial in terms of interpretability but is restrictive for real-life applications since it cannot discover hidden nonlinearities and interactions within the data. This study uses discrete-time survival models with embedded neural networks as estimators of time to default. This provides flexibility to express nonlinearities and interactions between variables and hence allows for models with better overall model fit. Additionally, the neural networks are used to estimate age–period–cohort (APC) models so that default risk can be decomposed into time components for loan age (maturity), origination (vintage), and environment (e.g., economic, operational, and social effects). These can be built as general models or as local APC models for specific customer segments. The local APC models reveal special conditions for different customer groups. The corresponding APC identification problem is solved by a combination of regularization and fitting the decomposed environment time risk component to macroeconomic data since the environmental risk is expected to have a strong relationship with macroeconomic conditions. Our approach is shown to be effective when tested on a large publicly available US mortgage dataset. This novel framework can be adapted by practitioners in the financial industry to improve modeling, estimation, and assessment of credit risk.

Consistency of banks' internal probability of default estimates: Empirical evidence from the COVID-19 crisis

The Basel III post-crisis reforms target the application of internal credit risk models for the estimation of the risk weighted assets of banks due to concerns about model risk. We use a unique dataset of 4.9 million probability of default (PD) estimates covering the January 2016 - June 2020 period sourced from 28 global banks to provide a deep insight into the comparability of model outputs. Our contribution is four-fold. Firstly, we confirm that there is a substantial variance in credit risk estimates. Secondly, we show that the level of PD variance is dependent on the entity type, industry, and location. Thirdly, we conclude that a considerable part of the variance is systematic, especially for credit risk estimates of funds. Finally, we illustrate the massive impact of the COVID-19 pandemic on the PD variance. The results highlight areas with relatively larger comparability issues, and they can be used by regulators to design more targeted policies.

Credit Risk Analysis Using Naive Bayes in Machine Learning

Abstract: A method of data analysis called machine learning automates the creation of analytical models. It is an area of artificial intelligence based on the idea that robots can learn from data, recognise patterns, and form opinions with little help from people. One of the most important and difficult tasks facing banks is the evaluation of credit risk. To prevent financial loss, accuracy is crucial in the classification of credit data. Lending money to people who need it is a major operation in the banking sector. The depositor bank receives the interest earned by the main borrowers in order to repay the principal borrowed from them. Financial risk management is increasingly focusing on credit risk analysis. The ability to estimate credit risk, monitor it, trust the model, and process loans well are essential for decision-making and transparency. In this study, we develop binary classifiers based on machine learning and deep learning models on actual data to forecast the likelihood of loan default. In the aforementioned research, we examine various methods for credit risk analysis that are used to the assessment of credit risk datasets.

Ellipsoidal buffered area under the curve maximization model with variable selection in credit risk estimation

Ellipsoidal buffered area under the curve maximization model with variable selection in credit risk estimation

Calibration alternatives to logistic regression and their potential for transferring the statistical dispersion of discriminatory power into uncertainties in probabilities of default

The transformation of credit scores into probabilities of default plays an important role in credit risk estimation. The linear logistic regression has developed into a standard calibration approach in the banking sector. With the advent of machine learning techniques in the discriminatory phase of credit risk models, however, the standard calibration approach is currently under scrutiny again. In particular, the assumptions behind the linear logistic regression provide critics with a target. The previous literature has converted the calibration problem into a regression task without any loss of generality. In this paper, we draw on recent academic results in order to suggest two new single-parameter families of differentiable functions as candidates for this regression. The derivation of these two families of differentiable functions is based on the maximum entropy principle, and thus it relies on a minimum number of assumptions. We compare the performance of four calibration approaches on a realworld data set and find that one of the new single-parameter families outperforms the linear logistic regression. Further, we develop an approach to quantify that part of the general estimation error of probabilities of default which stems from the statistical dispersion of the discriminatory power.

Integration of factor analysis and Tsukamoto’s fuzzy logic method for quality control of credit provisions in rural banks

Giving credit to debtors can pose a default risk. This risk arises because of an error in analyzing the credit risk rate of the debtor. Therefore, this study aims to design a framework for analyzing the credit risk rate of debtors so that the default risk can be reduced. This framework is created using the integration of factor analysis and Tsukamoto’s fuzzy logic method. This integration method can group many credit assessment variables into several decisive factors. In addition, the integration method can estimate credit risk rate firmly based on the α-predicate of each basic rule. This analytical framework is simulated on credit application data at a Rural Bank, in Indonesia. The simulation results show that there are three factors and one variable to measure the credit risk rate, namely: factor 1 represents repayment capacity, business length, working capital, and liquidity value; factor 2 represents the age and the difference between the granted and the proposed loan amount; factor 3 represents the stay length, character, and credit history; and one variable represents a dependent number. This research is expected to help credit institutions measure the credit risk rate in making credit decisions for prospective debtors.

High and Low Credit Risk in SME Portfolios: Evidence from Regulatory Risk Grade Dissemination

Purpose: SME sector credit risk has received attention in research from several dimensions of the financial system. SME sector’s funding is mainly supplied by financial institutions and SME sector is both diversified and large sector in both well developed and less developed economies. Specific research on assessing SME as Financial Institution’s (FI’s) individual counterparties and SMEs as portfolios have developed from a theoretical and empirical perspective. To supplement current research on the area, we approach SME risk from perspective of FIs own risk assessments and compare it to how SME risk rating and measurement compares to other counterparties. Design/methodology/approach: We use published risk rating data from large financial institutions in Europe including globally operating FIs and compare shares of credits in different risk grades and overall portfolio risks within an institution’s own risk classification system and risk measurement system. The data consists of 89 comparable portfolios with over 25 million credits. Findings: Our results show that comparison to households and large corporates originates from higher default rate estimates for SME, which shows as smaller share of credits in the investment grade. SME risk is further raised as even within speculative grades SME’s receives higher default estimates in comparison to households and large corporates. An equally notable finding is that the other relevant parameter for risk calculation, loss given default (LGD), does not differ between SME’s and other counterparties. A part of SME credits is found to be in a low-risk regime in portfolio credit risk estimation. Research limitations/implications: Coverage and detail of data restricts to a specified geographical coverage and aggregated data on SME-companies is not as exact as unit level data. The data represents mostly European institutions as it is collected from institutions which have a head quarter in Europe and are applying Basel regulation in a single rule book environment for banking regulation. In a global scopethere may be differences between jurisdictions or between geographical areas. Data published by institutions is an aggregated data on a rating grade level a and not on a unit level data that institutions have for their exact calculations. Comparison methods for SME sector are selected accordingly so that methods apply to class level instead of unit level data. Originality/value: Higher capital requirements for SME’s may restrict the price and availability of finance. According to our results there can be separated a low-risk SME finance without higher capital requirements compared to peers. Results may also be used to support counterparty level default risk model results showing higher risk for SME’s which can be seen in smaller shares of investment grade credits and in a higher default rate for speculative grade credits.

What can we learn about credit risk from debt valuation adjustments?

Motivated by the debate about the introduction of the fair value option for (financial) liabilities (FVOL) and the requirement to recognize and separately disclose in financial statements debt valuation adjustments (DVAs), this study explores what we can learn about a firm’s credit risk from DVAs. Using a sample of US bank holding companies that elect the FVOL, we show that DVAs generally cannot be explained by the same factors that explain contemporaneous changes in bank’s credit quality. We further find that DVAs can explain future changes in credit risk when the fair value of liabilities is based on managerial inputs (Level 3). Overall our results suggest that managers have an information advantage in estimating credit risk and that DVAs provide inside information to the market.

Structural estimation of counterparty credit risk under recovery risk

Counterparty Credit Risk (CCR) represents one of the major sources of uncertainty in many financial contracts. The role of credit value adjustment (CVA) is, in fact, that of rewarding the parties for the exposure to such risk. A key driver of CVA is the recovery risk, generated by the variability of recovery rates. In this paper, we develop a framework to assess the CCR accounting for the recovery risk that arises from the introduction of stochastic recovery rates. Adopting the structural model for the time to default that exploits a time-changed Lévy process for the risk driver of the equity value, we provide a complete picture to monitor the CCR and gauge the effects of the stochastic recovery rates. The model extracts information on the creditworthiness of the parties in the OTC contract combining Fourier Cosine Expansion and Monte Carlo simulations methods to price CDS spreads, the related underlying, and to retrieve the default barrier. We apply the model proposed to a business case analyzing the CCR of two parties involved in the OTC contract with underlying energy commodities. Low average recovery rates reveal to be associated with high implied volatility and depart from the fixed value of 40%, especially during periods of market distress.

Bank-sourced transition matrixes: are banks’ internal credit risk estimates Markovian?

This study explores banks’ internal credit risk estimates and the associated banksourced transition matrixes. We empirically test the widely used Markovian property and time homogeneity assumptions at an unprecedented scale. Our unique data set consists of internal probability of default estimates from 12 global banks that employ an advanced internal ratings-based approach, covering monthly observations on 20,000 corporates over the period 2015–18. The results indicate that internal credit risk estimates do not satisfy the two assumptions, showing evidence of both path dependency and time heterogeneity even within the period of economic expansion. Banks tend to revert their rating actions, which contradicts previous findings based on data from credit rating agencies. This has significant practical implications through bank-sourced credit transition matrixes, which are becoming increasingly important as regulators begin to use more detailed credit risk data sets (eg, analytical credit data sets by the European Central Bank) with potential applications in stress testing.

Multi-heuristic scheduling methods for workflow in credit cloud

In the credit cloud, credit services are sold to applications for credit computing, credit fusion and credit risk estimates. Plenty of services with different performance for the same task may have different execution time and charged by various ways. The users have specific requirements for the workflow completion time or cost. Hence, to meet the user’s satisfaction is an important challenge. In this paper, we propose heuristic scheduling methods for credit workflow with total cost minimization, and the deadline should be satisfied. The problem can be divided into two sub-problems, task-mode mapping and task tabling on renting service instances. For the task-mode mapping problem, a recursive heuristic method is constructed to select appropriate service for each task of the workflow. Then another heuristic algorithm based is established to get a final schema with deadline constraint. We discussed the service instance rented in shareable manner and compared with un-shareable manner. Three renting strategies are discussed in detail. Experimental results show the effectiveness and efficiency of the proposed algorithm.

SHAP and LIME: An Evaluation of Discriminative Power in Credit Risk.

In credit risk estimation, the most important element is obtaining a probability of default as close as possible to the effective risk. This effort quickly prompted new, powerful algorithms that reach a far higher accuracy, but at the cost of losing intelligibility, such as Gradient Boosting or ensemble methods. These models are usually referred to as “black-boxes”, implying that you know the inputs and the output, but there is little way to understand what is going on under the hood. As a response to that, we have seen several different Explainable AI models flourish in recent years, with the aim of letting the user see why the black-box gave a certain output. In this context, we evaluate two very popular eXplainable AI (XAI) models in their ability to discriminate observations into groups, through the application of both unsupervised and predictive modeling to the weights these XAI models assign to features locally. The evaluation is carried out on real Small and Medium Enterprises data, obtained from official italian repositories, and may form the basis for the employment of such XAI models for post-processing features extraction.

Predicting loss given default using post-default information

The loss given default (LGD) is an essential component for estimating credit risk according to the international regulatory Basel Accord. Traditionally, LGD models are built based on the characteristics of the loan and the borrower prior to default, a practice which fails to consider the post-default information revealed during the repayment process. We start by uncovering a predictive post-default variable (i.e., a flag that indicates whether the defaulted borrower had cooperated with a debt-settlement company) in the defaulted data from the online lending platform. We then propose a stratified modelling framework to incorporate this variable into LGD prediction. The experimental results demonstrate that LGD prediction is significantly improved by the inclusion of post-default information under the proposed stratified modelling framework. We further show that the predictive performance of the proposed models is robust for the choice of training set and input variables. Our results imply the importance of using post-default information in LGD prediction.

A comparison of the performance of alternative Machine Learning algorithms on a credit risk dataset

We compare alternative methodologies for Credit Risk estimation. Here we introduce a Bayesian Model Averaging (BMA) of Logistic Regressions for Probability of Default modeling where the model space is sampled via Markov Chain Monte Carlo methods making use of the approximation of the posterior probability introduced in [Schwarz, 1978] and used in the context of sparse problems in [Sala-i Martin et al., 2004], [Gross and Poblacion, 2004] and [Torresetti, 2021]. We will compare this methodology against some popular Machine Learning ensemble methods for regression trees: Bagging (Bootstrap Aggregating) and Boosting (Adaptive Boosting, Gradient Boosting). The results on the credit risk dataset at our disposal show how Gradient Boosting proves to be the top performer but with a slow convergence speed given the cooling coecient that was used to make it less greedy. As not too distant seconds we see the performance of BMA and Adaptive Boosting where in particular BMA proves to have a faster convergence speed among the specific versions of the algorithms implemented. Finally a simple average of estimates as in Bagging proves to have the lowest performance of the ensemble methods considered.

The review of the open challenges in the IRB loan portfolio credit risk modeling

The Basel Committee on Banking Supervision finalized the Basel III accord in the December 2017 and launched the set of its standards – the Basel Framework – in December 2019. Both documents allow bank to use mathematical models for the credit risk estimation. There are quantitative and qualitative requirements for models to be allowed for use in the prudential regulation of banks. The approach is called an Internal-Ratings-Based one (IRB). This paper aims at discussing a set of issues related to IRB credit risk modeling and such model estimates use. Those issues include data pooling in the credit registries, applying copula-discriminant analysis, validating the borrower concentration per grade, assigning the hybrid credit rating, use of model estimates when voting at the credit committee, estimate of the ultimate credit risk-taking by banks.

Determinants of defaults on P2P lending platforms in China

Chinese P2P lending platforms have an astonishing default rate of 87.2% based on data available in 2019, which indicates the seriousness of the problem this industry faces. Insufficient regulation has resulted in generation of risky services, such as margin finance in 2015 for stock markets and zero down-payment mortgages in 2016 for real estate buyers. Such services are prone to resulting in dramatic losses to investors with the following potential causes: adverse selection caused by information asymmetry of the P2P platform operators, lack of financial knowledge or expertise of the investors, insufficient regulation on P2P platforms, and changes in policies related to stock and real estate markets.Does “Too Big to Fail” (TBTF) apply to Peer-to-peer (P2P) lending platforms in China? Are platforms with large Transaction Volumes or large Total Loan Size safer? This paper aims at answering such questions and identifies other factors that have led to defaults of such P2P lending platforms. In order to generate a more robust model that covers a wider range of factors, we adopt two approaches: Failure Prediction Model and Decomposition Methods. The reason why we have adopted this approach was that the factors that caused failure of such P2P platforms were very complicated and varied in nature. Failure Prediction Model is mainly based on the disclosed transaction or financial data. However, Decomposition Method aims at identification of sources of problem from different groups of factors, such as internal operations and external business environment. Also, another contribution of our research is that the data sets that we have used are from three different sources, which provide us a more comprehensive picture about the risks faced by such P2P lending firms.This study provides evidence that the bond yield, which reflects the liquidity level of the market, is the major reason for P2P platforms’ failure during the research period. Our results also indicate that apart from the size-related factors one also needs to answer the question - “Were platforms with larger Transaction Volumes and Total Loan Sizes safer?” The style factors like the interest rate and ownership, the timing factors like loan maturity and bond yield, and the selectivity factors like popularity are found to be even more significant. Hopefully, our findings could contribute academically to the existing literature in the credit risk estimation of online P2P lending platforms and provide practical support for development of a more effective risk monitoring system for online P2P lending platforms.

Bank-sourced credit transition matrices: Estimation and characteristics

This study proposes and analyses a novel alternative to credit transition matrices (CTMs) developed by credit rating agencies - bank-sourced CTMs. It provides a unique insight into estimation of bank-sourced CTMs by assessing the extent to which the CTMs depend on the characteristics of the underlying credit risk datasets and the aggregation method and outlines that the choice of aggregation approach has a substantial effect on credit risk model results. Further, we show that bank-sourced CTMs are more dynamic than those of credit rating agencies, with higher off-diagonal transition rates and higher propensity to upgrade. Finally, we create a set of industry-specific CTMs, otherwise unobtainable due to the data sparsity faced by credit rating agencies, and highlight the implications of their differences, signalling the existence of industry-specific business cycles. The study uses a unique and large dataset of internal credit risk estimates from 24 global banks covering monthly observations on more than 26,000 large corporates and employs large-scale Monte Carlo simulations. This approach can be replicated by regulators (e.g., data collected by the European Central Bank in the AnaCredit project) and used by organisations aiming to improve their credit risk models.

Die Hard: Probability of Default and Soft Information

The research aims to verify whether the credit risk of small and medium-sized enterprises can be estimated more accurately using qualitative variables together with financial information from reports. In our paper, we select qualitative variables within the conceptual framework of the balanced scorecard to assess the credit quality of Italian companies of various sizes, from micro to medium. Data were collected to estimate the company’s resilience following the shock of the financial crisis of 2007–2008. The analysis based on customer size, processes, knowledge, and corporate finance, synthesized with balanced scorecard methodology, allows us to estimate the resilience of companies in a period of crisis. The research highlights the important contribution of qualitative variables for the estimation of credit risk. The implications concern both financial intermediaries and their supervisory functions, and regulators for rating models based on soft forward and countercyclical variables.

Comparison study of two-step LGD estimation model with probability machines

Accurate estimation of loss given default is necessary to estimating credit risk. Due to the bi-modal nature of LGD, the two-step LGD estimation model is a promising method for LGD estimation. This study improves the first model in the two-step LGD estimation model using probability machines (random forest, k-nearest neighbors, bagged nearest neighbors, and support vector machines). Furthermore, we compare the predictive performance of each model with traditional logistic regression models. This study confirms that random forest is the best model for developing the first model in the two-step LGD estimation model.