Recent literature implements machine learning techniques to assess corporate credit rating based on financial statement reports. In this work, we analyze the performance of four neural network architectures (MLP, CNN, CNN2D, LSTM) in predicting corporate credit rating as issued by Standard and Poor's. We analyze companies from the energy, financial and healthcare sectors in US. The goal of the analysis is to improve application of machine learning algorithms to credit assessment. To this end, we focus on three questions. First, we investigate if the algorithms perform better when using a selected subset of features, or if it is better to allow the algorithms to select features themselves. Second, is the temporal aspect inherent in financial data important for the results obtained by a machine learning algorithm? Third, is there a particular neural network architecture that consistently outperforms others with respect to input features, sectors and holdout set? We create several case studies to answer these questions and analyze the results using ANOVA and multiple comparison testing procedure.
Credit ratings are one of the primary keys that reflect the level of riskiness and reliability of corporations to meet their financial obligations. Rating agencies tend to take extended periods of time to provide new ratings and update older ones. Therefore, credit scoring assessments using artificial intelligence has gained a lot of interest in recent years. Successful machine learning methods can provide rapid analysis of credit scores while updating older ones on a daily time scale. Related studies have shown that neural networks and support vector machines outperform other techniques by pr…
Erasmus University Rotterdam（伊拉斯姆斯大学鹿特丹分校）；ERIM（伊拉斯姆斯管理研究所）；Tinbergen Institute（丁伯根研究所）；CREATES（高级经济时间序列研究中心）
Sovereign credit ratings summarize the creditworthiness of countries. These ratings have a large influence on the economy and the yields at which governments can issue new debt. This paper investigates the use of a Multilayer Perceptron (MLP), Classification and Regression Trees (CART), Support Vector Machines (SVM), Naïve Bayes (NB), and an Ordered Logit (OL) model for the prediction of sovereign credit ratings. We show that MLP is best suited for predicting sovereign credit ratings, with a random cross-validated accuracy of 68%, followed by CART (59%), SVM (41%), NB (38%), and OL (33%). Inve…
Central University of Finance and Economics（中央财经大学）；Institute of Automation, Chinese Academy of Sciences（中国科学院自动化研究所）；University of Chinese Academy of Sciences（中国科学院大学）
Due to the powerful learning ability on high-rank and non-linear features, deep neural networks (DNNs) are being applied to data mining and machine learning in various fields, and exhibit higher discrimination performance than conventional methods. However, the applications based on DNNs are rare in enterprise credit rating tasks because most of DNNs employ the "end-to-end" learning paradigm, which outputs the high-rank representations of objects and predictive results without any explanations. Thus, users in the financial industry cannot understand how these high-rank representations are gene…
Enhancing Credit Risk Prediction: A Multi-stage Ensemble Pipeline
Division of International Business and Technology Studies, A.R. Sánchez Jr. School of Business, Texas A&M International University（德克萨斯A&M国际大学A.R. Sánchez Jr.商学院国际商务与技术研究部）；Dept. of Management and Marketing, Angelo State University（安吉洛州立大学管理与市场营销系）；Department of Management and Marketing, Southern University and A&M College（南方大学与农工学院管理与市场营销系）；Department of METE, Firat University（菲拉特大学METE系）；Department of Accounting, Western Washington University（西华盛顿大学会计系）
Credit rating is an analysis of the credit risks associated with a corporation, which reflect the level of the riskiness and reliability in investing. There have emerged many studies that implement machine learning techniques to deal with corporate credit rating. However, the ability of these models is limited by enormous amounts of data from financial statement reports. In this work, we analyze the performance of traditional machine learning models in predicting corporate credit rating. For utilizing the powerful convolutional neural networks and enormous financial data, we propose a novel en…
In Artificial Intelligence, interpreting the results of a Machine Learning technique often termed as a black box is a difficult task. A counterfactual explanation of a particular "black box" attempts to find the smallest change to the input values that modifies the prediction to a particular output, other than the original one. In this work we formulate the problem of finding a counterfactual explanation as an optimization problem. We propose a new "sparsity algorithm" which solves the optimization problem, while also maximizing the sparsity of the counterfactual explanation. We apply the spar…
We explicitly test if the reliability of credit ratings depends on the total number of admissible states. We analyse open access credit rating data and show that the effect of the number of states in the dynamical properties of ratings change with time, thus giving supportive evidence that the ideal number of admissible states changes with time. We use matrix estimation methods that explicitly assume the hypothesis needed for the process to be a valid rating process. By comparing with the likelihood maximization method of matrix estimation, we quantify the "likelihood-loss" of assuming that th…
Risk management is an important practice in the banking industry. In this paper we develop a new methodology to estimate and predict the probability of default (PD) based on the rating transition matrices, which relates the rating transition matrices to the macroeconomic variables. Our method can overcome the shortcomings of the framework of Belkin et al. (1998), and is especially useful in predicting the PD and doing stress testing. Simulation is conducted at the end, which shows that our method can provide more accurate estimate than that obtained by the method of Belkin et al. (1998).
We compare observed corporate cumulative default probabilities to those calculated using a stochastic model based on an extension of the work of Black and Cox and find that corporations default as if via diffusive dynamics. The model, based on a contingent-claims analysis of corporate capital structure, is easily calibrated with readily available historical default probabilities and fits observed default data published by Standard and Poor's. Applying this model to the Standard and Poor's default data we find that the difference in default behavior between credit ratings can be explained large…
We present a continuous-time maximum likelihood estimation methodology for credit rating transition probabilities, taking into account the presence of censored data. We perform rolling estimates of the transition matrices with exponential time weighting with varying horizons and discuss the underlying dynamics of transition generator matrices in the long-term and short-term estimation horizons.
Aggregate and systemic risk in complex systems are emergent phenomena depending on two properties: the idiosyncratic risks of the elements and the topology of the network of interactions among them. While a significant attention has been given to aggregate risk assessment and risk propagation once the above two properties are given, less is known about how the risk is distributed in the network and its relations with the topology. We study this problem by investigating a large proprietary dataset of payments among 2.4M Italian firms, whose credit risk rating is known. We document significant c…
In this paper we discuss the issue of computation of the bilateral credit valuation adjustment (CVA) under rating triggers, and in presence of ratings-linked margin agreements. Specifically, we consider collateralized OTC contracts, that are subject to rating triggers, between two parties -- an investor and a counterparty. Moreover, we model the margin process as a functional of the credit ratings of the counterparty and the investor. We employ a Markovian approach for modeling of the rating transitions of the two parties to the contract. In this framework, we derive the representation for bil…
We compare two models of corporate default by calculating the Jeffreys-Kullback-Leibler divergence between their predicted default probabilities when asset correlations are either high or low. Our main results show that the divergence between the two models increases in highly correlated, volatile, and large markets, but that it is closer to zero in small markets, when asset correlations are low and firms are highly leveraged. These findings suggest that during periods of financial instability the single-and multi-factor models of corporate default will generate increasingly inconsistent predi…
We introduce a simple approach for testing the reliability of homogeneous generators and the Markov property of the stochastic processes underlying empirical time series of credit ratings. We analyze open access data provided by Moody's and show that the validity of these assumptions - existence of a homogeneous generator and Markovianity - is not always guaranteed. Our analysis is based on a comparison between empirical transition matrices aggregated over fixed time windows and candidate transition matrices generated from measurements taken over shorter periods. Ratings are widely used in cre…
Small Medium-sized Enterprises (SMEs) face many obstacles when they try to access credit market. These obstacles are increased if the SMEs are innovative. In this case, financial data are insufficient or even not reliable. Thus, when building a judgemental rating model, mainly based on qualitative criteria (soft information), it is very important to finance SMEs' activities. Until now, there isn't a multicriteria credit risk model based on soft information for innovative SMEs. In this paper, we try to fill this gap by presenting a multicriteria credit risk model, specifically, ELECTRE-TRI. To …
We present two methodologies on the estimation of rating transition probabilities within Markov and non-Markov frameworks. We first estimate a continuous-time Markov chain using discrete (missing) data and derive a simpler expression for the Fisher information matrix, reducing the computational time needed for the Wald confidence interval by a factor of a half. We provide an efficient procedure for transferring such uncertainties from the generator matrix of the Markov chain to the corresponding rating migration probabilities and, crucially, default probabilities. For our second contribution, …
We analyse the effectiveness of modern deep learning techniques in predicting credit ratings over a universe of thousands of global corporate entities obligations when compared to most popular, traditional machine-learning approaches such as linear models and tree-based classifiers. Our results show a adequate accuracy over different rating classes when applying categorical embeddings to artificial neural networks (ANN) architectures.
We propose a Markov chain model for credit rating changes. We do not use any distributional assumptions on the asset values of the rated companies but directly model the rating transitions process. The parameters of the model are estimated by a maximum likelihood approach using historical rating transitions and heuristic global optimization techniques. We benchmark the model against a GLMM model in the context of bond portfolio risk management. The proposed model yields stronger dependencies and higher risks than the GLMM model. As a result, the risk optimal portfolios are more conservative th…
We consider the problem of constructing an appropriate multivariate model for the study of the counterparty credit risk in credit rating migration problem. For this financial problem different multivariate Markov chain models were proposed. However the markovian assumption may be inappropriate for the study of the dynamic of credit ratings which typically show non markovian like behaviour. In this paper we develop a semi-Markov approach to the study of the counterparty credit risk by defining a new multivariate semi-Markov chain model. Methods are given for computing the transition probabiliti…
In this work we build a stack of machine learning models aimed at composing a state-of-the-art credit rating and default prediction system, obtaining excellent out-of-sample performances. Our approach is an excursion through the most recent ML / AI concepts, starting from natural language processes (NLP) applied to economic sectors' (textual) descriptions using embedding and autoencoders (AE), going through the classification of defaultable firms on the base of a wide range of economic features using gradient boosting machines (GBM) and calibrating their probabilities paying due attention to t…
The use of CVA to cover credit risk is widely spread, but has its limitations. Namely, dealers face the problem of the illiquidity of instruments used for hedging it, hence forced to warehouse credit risk. As a result, dealers tend to offer a limited OTC derivatives market to highly risky counterparties. Consequently, those highly risky entities rarely have access to hedging services precisely when they need them most. In this paper we propose a method to overcome this limitation. We propose to extend the CVA risk-neutral framework to compute an initial margin (IM) specific to each counterpart…
The existence of asymmetric information has always been a major concern for financial institutions. Financial intermediaries such as commercial banks need to study the quality of potential borrowers in order to make their decision on corporate loans. Classical methods model the default probability by financial ratios using the logistic regression. As one of the major commercial banks in France, we have access to the the account activities of corporate clients. We show that this transactional data outperforms classical financial ratios in predicting the default event. As the new data reflects t…
Corporate defaults may be triggered by some major market news or events such as financial crises or collapses of major banks or financial institutions. With a view to develop a more realistic model for credit risk analysis, we introduce a new type of reduced-form intensity-based model that can incorporate the impacts of both observable "trigger" events and economic environment on corporate defaults. The key idea of the model is to augment a Cox process with trigger events. Both single-default and multiple-default cases are considered in this paper. In the former case, a simple expression for t…
In this paper, we propose a methodology based on piece-wise homogeneous Markov chain for credit ratings and a multivariate model of the credit spreads to evaluate the financial risk in European Union (EU). Two main aspects are considered: how the financial risk is distributed among the European countries and how large is the value of the total risk. The first aspect is evaluated by means of the expected value of a dynamic entropy measure. The second one is solved by computing the evolution of the total credit spread over time. Moreover, the covariance between countries' total spread allows und…
Banks and financial institutions all over the world manage portfolios containing tens of thousands of customers. Not all customers are high credit-worthy, and many possess varying degrees of risk to the Bank or financial institutions that lend money to these customers. Hence assessment of credit risk is paramount in the field of credit risk management. This paper discusses the use of Bayesian principles and simulation-techniques to estimate and calibrate the default probability of credit ratings. The methodology is a two-phase approach where, in the first phase, a posterior density of default …
The Basel II internal ratings-based (IRB) approach to capital adequacy for credit risk implements an asymptotic single risk factor (ASRF) model. Measurements from the ASRF model of the prevailing state of Australia's economy and the level of capitalisation of its banking sector find general agreement with macroeconomic indicators, financial statistics and external credit ratings. However, given the range of economic conditions, from mild contraction to moderate expansion, experienced in Australia since the implementation of Basel II, we cannot attest to the validity of the model specification …
Interpretable LLMs for Credit Risk: A Systematic Review and Taxonomy
School of Electronic Engineering and Computer Science, Queen Mary University of London（伦敦大学电子工程与计算机科学学院）；Institute for Data Science and Artificial Intelligence, Boğaziçi University（博雅大学数据科学与人工智能研究所）；College of Science Computer Science Department, Kuwait University（科威特大学科学学院计算机科学系）
Inspired by the bankruptcy of Lehman Brothers and its consequences on the global financial system, we develop a simple model in which the Lehman default event is quantified as having an almost immediate effect in worsening the credit worthiness of all financial institutions in the economic network. In our stylized description, all properties of a given firm are captured by its effective credit rating, which follows a simple dynamics of co-evolution with the credit ratings of the other firms in our economic network. The dynamics resembles the evolution of Potts spin-glass with external global f…
Università di Bologna（博洛尼亚大学）；Bergische Universität Wuppertal（伍珀塔尔大学）
In this paper, we model the rating process of an entity by using a geometrical approach. We model rating transitions as an SDE on a Lie group. Specifically, we focus on calibrating the model to both historical data (rating transition matrices) and market data (CDS quotes) and compare the most popular choices of changes of measure to switch from the historical probability to the risk-neutral one. For this, we show how the classical Girsanov theorem can be applied in the Lie group setting. Moreover, we overcome some of the imperfections of rating matrices published by rating agencies, which are …
The importance of adequately modeling credit risk has once again been highlighted in the recent financial crisis. Defaults tend to cluster around times of economic stress due to poor macro-economic conditions, {\em but also} by directly triggering each other through contagion. Although credit default swaps have radically altered the dynamics of contagion for more than a decade, models quantifying their impact on systemic risk are still missing. Here, we examine contagion through credit default swaps in a stylized economic network of corporates and financial institutions. We analyse such a syst…
Credit risk management, the practice of mitigating losses by understanding the adequacy of a borrower's capital and loan loss reserves, has long been imperative to any financial institution's long-term sustainability and growth. MassMutual is no exception. The company is keen on effectively monitoring downgrade risk, or the risk associated with the event when credit rating of a company deteriorates. Current work in downgrade risk modeling depends on multiple variations of quantitative measures provided by third-party rating agencies and risk management consultancy companies. As these structure…
Bond rating Transition Probability Matrices (TPMs) are built over a one-year time-frame and for many practical purposes, like the assessment of risk in portfolios or the computation of banking Capital Requirements (e.g. the new IFRS 9 regulation), one needs to compute the TPM and probabilities of default over a smaller time interval. In the context of continuous time Markov chains (CTMC) several deterministic and statistical algorithms have been proposed to estimate the generator matrix. We focus on the Expectation-Maximization (EM) algorithm by Bladt and Sorensen (2005) for a CTMC with an abs…
Diffusion in a linear potential in the presence of position-dependent killing is used to mimic a default process. Different assumptions regarding transport coefficients, initial conditions, and elasticity of the killing measure lead to diverse models of bankruptcy. One "stylized fact" is fundamental for our consideration: empirically default is a rather rare event, especially in the investment grade categories of credit ratings. Hence, the action of killing may be considered as a small parameter. In a number of special cases we derive closed-form expressions for the entire term structure of th…
Enhancing ML Models Interpretability for Credit Scoring
Dept. of Computer Science, Delft University of Technology（代尔夫特理工大学计算机科学系）；Dept. of Applied Mathematics, Delft University of Technology（代尔夫特理工大学应用数学系）；FF Quant Advisory B.V.（FF量化咨询公司）
