Modelos de credit score são usualmente desenvolvidos somente com informações dos proponentes aceitos. Neste trabalho foram consideradas estratégias que podem ser utilizadas para o desenvolvimento de modelos de credit score com a inclusão das informações dos rejeitados. Foram avaliadas as seguintes técnicas de inferência de rejeitados: classificação dos rejeitados como clientes Maus, parcelamento, dados aumentados, uso de informações de mercado e ainda a estratégia de aceitar proponentes rejeitados para acompanhamento e desenvolvimento de novos modelos de risco de crédito. Para a avaliação e comparação dos modelos foram utilizadas as medidas de desempenho: estatística de Kolmogorov-Smirnov (KS), área sob a curva de Lorentz (ROC), área entre as curvas de distribuição acumulada dos escores (AEC), diferença entre as taxas de inadimplência nos intervalos do escore definidos pelos decis e coeficiente de Gini. Concluiu-se que dentre as quatro primeiras técnicas avaliadas, o uso de informaçõoes de mercado foi a que apresentou melhor desempenho. Quanto à estratégia de aceitar proponentes rejeitados, observou-se que há um ganho em relação ao modelo ajustado só com base nos proponentes aceitos. / Credit scoring models are usually built using only information of accepted applicants. This text considered strategies that can be used to develop credit score models with inclusion of the information of the rejects. We evaluated the techniques of reject inference: classification of rejected customers as bad, parceling, augmentation, use of market information and the strategy of accepting rejected proponents for monitoring and developing new models of credit risk. For the evaluation and comparison between models were used performance measures: Kolmogorov-Smirnov statistics (KS), the area under the Lorentz Curve (ROC), area between cumulative distribution curves of the scores (AEC), difference among the delinquency rate in the score buckets based on deciles (DTI) and the Gini coefficient. We concluded that among the first four techniques evaluated, the fourth (use of market information) had the best performance. For the strategy to accept rejected bidders, it was observed that there is a gain in relation to the model that uses only information of accepted applicants.

Modelos de credit score são usualmente desenvolvidos somente com informações dos proponentes aceitos. Neste trabalho foram consideradas estratégias que podem ser utilizadas para o desenvolvimento de modelos de credit score com a inclusão das informações dos rejeitados. Foram avaliadas as seguintes técnicas de inferência de rejeitados: classificação dos rejeitados como clientes Maus, parcelamento, dados aumentados, uso de informações de mercado e ainda a estratégia de aceitar proponentes rejeitados para acompanhamento e desenvolvimento de novos modelos de risco de crédito. Para a avaliação e comparação dos modelos foram utilizadas as medidas de desempenho: estatística de Kolmogorov-Smirnov (KS), área sob a curva de Lorentz (ROC), área entre as curvas de distribuição acumulada dos escores (AEC), diferença entre as taxas de inadimplência nos intervalos do escore definidos pelos decis e coeficiente de Gini. Concluiu-se que dentre as quatro primeiras técnicas avaliadas, o uso de informaçõoes de mercado foi a que apresentou melhor desempenho. Quanto à estratégia de aceitar proponentes rejeitados, observou-se que há um ganho em relação ao modelo ajustado só com base nos proponentes aceitos. / Credit scoring models are usually built using only information of accepted applicants. This text considered strategies that can be used to develop credit score models with inclusion of the information of the rejects. We evaluated the techniques of reject inference: classification of rejected customers as bad, parceling, augmentation, use of market information and the strategy of accepting rejected proponents for monitoring and developing new models of credit risk. For the evaluation and comparison between models were used performance measures: Kolmogorov-Smirnov statistics (KS), the area under the Lorentz Curve (ROC), area between cumulative distribution curves of the scores (AEC), difference among the delinquency rate in the score buckets based on deciles (DTI) and the Gini coefficient. We concluded that among the first four techniques evaluated, the fourth (use of market information) had the best performance. For the strategy to accept rejected bidders, it was observed that there is a gain in relation to the model that uses only information of accepted applicants.

Le credit scoring est généralement considéré comme une méthode d’évaluation du niveau du risque associé à un dossier de crédit potentiel. Cette méthode implique l'utilisation de différentes techniques statistiques pour aboutir à un modèle de scoring basé sur les caractéristiques du client.Le modèle de scoring estime le risque de crédit en prévoyant la solvabilité du demandeur de crédit. Les institutions financières utilisent ce modèle pour estimer la probabilité de défaut qui va être utilisée pour affecter chaque client à la catégorie qui lui correspond le mieux: bon payeur ou mauvais payeur. Les seules données disponibles pour construire le modèle de scoring sont les dossiers acceptés dont la variable à prédire est connue. Ce modèle ne tient pas compte des demandeurs de crédit rejetés dès le départ ce qui implique qu'on ne pourra pas estimer leurs probabilités de défaut, ce qui engendre un biais de sélection causé par la non-représentativité de l'échantillon. Nous essayons dans ce travail en utilisant l'inférence des refusés de remédier à ce biais, par la réintégration des dossiers refusés dans le processus d'octroi de crédit. Nous utilisons et comparons différentes méthodes de traitement des refusés classiques et semi supervisées, nous adaptons certaines à notre problème et montrons sur un jeu de données réel, en utilisant les courbes ROC confirmé par simulation, que les méthodes semi-supervisé donnent de bons résultats qui sont meilleurs que ceux des méthodes classiques. / Credit scoring is generally considered as a method of evaluation of a risk associated with a potential loan applicant. This method involves the use of different statistical techniques to determine a scoring model. Like any statistical model, scoring model is based on historical data to help predict the creditworthiness of applicants. Financial institutions use this model to assign each applicant to the appropriate category : Good payer or Bad payer. The only data used to build the scoring model are related to the accepted applicants in which the predicted variable is known. The method has the drawback of not estimating the probability of default for refused applicants which means that the results are biased when the model is build on only the accepted data set. We try, in this work using the reject inference, to solve the problem of selection bias, by reintegrate reject applicants in the process of granting credit. We use and compare different methods of reject inference, classical methods and semi supervised methods, we adapt some of them to our problem and show, on a real dataset, using ROC curves, that the semi-supervised methods give good results and are better than classical methods. We confirmed our results by simulation.

Alors que les modèles de credit scoring sont largement utilisés depuis plus de cinquante ans et sont considérés comme un outil indispensable dans la prise de décision dans d'innombrables institutions financières du monde entier, la littérature et les empiriques disponibles sur ce sujet restent encore très limitées. Notre objectif est de combler cette lacune en présentant une analyse approfondie sur les modèles de credit scoring et le processus de prise de décision d’octroi de crédit, avec diverses applications sur des données réelles et extensives provenant de différents pays. Notre thèse comporte trois chapitres. Chapitre 1 commence par présenter le processus de développement d’un modèle de credit scoring, et fournit une application sur des données réelles d'une banque de détail basée en France. Visant à donner de nouvelles perspectives sur les pays émergents, Chapitre 2 analyse le marché du crédit à la consommation en Chine et enquête sur l'utilisation des modèles de credit scoring dans un tel marché prometteur. Chapitre 3 va plus loin que la littérature méthodologique précédente et se concentre sur les différentes techniques d'inférence des refusés qui peuvent corriger le biais de sélection lors de la construction d'un modèle de crédit scoring basé uniquement sur les dossiers acceptés. Ces chapitres présentent les différents aspects du crédit scoring, pour lesquels les principales problématiques de credit scoring seront traitées. / While credit scoring has been broadly used for more than fifty years and continued to be a great support on decision-making in countless businesses around the world, the amount of literature, especially empirical studies, available on this subject is still limited. Our aim in this thesis is to fill this gap by providing a profound analysis on credit scoring and credit decision processes, with various applications using real and extensive sets of data coming from different countries. The thesis is organized in three chapters. Chapter 1 starts by presenting the credit scoring development process, and provides an application to real data from a France-based retail bank. Aiming at providing new insights regarding emerging countries, Chapter 2 analyzes the Chinese consumer lending market and investigates the use of credit scoring in such a promising market. Chapter 3 goes further than the previous methodological literature and focuses on reject inference techniques which can be a way to address the bias when developing a credit-scoring model based solely on accepted applicants. These chapters provide a round tour on credit scoring, after which major issues in credit scoring are treated.

Nowadays, the use of credit scoring models in the financial sector is a common practice. Credit scoring plays an important role in profitability and transparency of lending business. Given the high credit volumes, even a small improvement of discriminatory and predictive power of a credit scoring model may provide a substantial additional profit. Scoring models are applied on the through-the-door population, however, for creating them or adjusting already existing credit rules, it is usual to use only the data corresponding to accepted applicants for which payment discipline can be observed. This discrepancy can lead to reject bias (or selection bias in general). Methods trying to eliminate or reduce this phenomenon are known by the term reject inference. In general, these methods try to assess the behavior of rejected applicants or to obtain an additional information about them. In the dissertation thesis, I dealt with the enlargement method which is based on a random acceptance of applicants that would have been rejected. This method is not only time consuming but also expensive. Therefore I looked for the ways how to reduce the cost of acquiring additional information about rejected applicants. As a result, I have proposed a modification which I called the enlargement method with sorting variable. It was validated on real bank database with two possible sorting variables and the results were compared with the original version of the method. It was shown that both tested approaches can reduce its cost while retaining the accuracy of the scoring models.

Machine learning plays a vital role in preventing financial losses within the banking industry, and still, a lot of state of the art and industry-standard approaches within the field neglect rejected customer information and the potential information that they hold to detect similar risk behavior.This thesis explores the possibility of including this information during training and utilizing transactional history through an LSTM to improve the detection of defaults.  The model is structured so an encoder is first trained with or without rejected customers. Virtual distances are then calculated in the embedding space between the accepted customers. These distances are used to create a graph where each node contains an LSTM network, and a GCN passes messages between connected nodes. The model is validated using two datasets, one public Taiwan dataset and one private Swedish one provided through the collaborative company. The Taiwan dataset used 8000 data points with a 50/50 split in labels. The Swedish dataset used 4644 with the same split.  Multiple metrics were used to validate the impact of the rejected customers and the impact of using time-series data instead of static features. For the encoder part, reconstruction error was used to measure the difference in performance. When creating the edges, the homogeny of the neighborhoods and if a node had a majority of the same labeled neighbors as itself were determining factors, and for the classifier, accuracy, f1-score, and confusion matrix were used to compare results. The results of the work show that the impact of rejected customers is minor when it comes to changes in predictive power. Regarding the effects of using time-series information instead of static features, we saw a comparative result to XGBoost on the Taiwan dataset and an improvement in the predictive power on the Swedish dataset. The results also show the importance of a well-defined virtual distance is critical to the classifier's performance.

In statistical methodology of analyzing data, two of the most commonly used techniques are classification and regression modeling. As scientific technology progresses rapidly, complex data often occurs and requires novel classification and regression modeling methodologies according to the data structure. In this dissertation, I mainly focus on developing a few approaches for analyzing the data with complex structures. Classification problems commonly occur in many areas such as biomedical, marketing, sociology and image recognition. Among various classification methods, linear classifiers have been widely used because of computational advantages, ease of implementation and interpretation compared with non-linear classifiers. Specifically, linear discriminant analysis (LDA) is one of the most important methods in the family of linear classifiers. For high dimensional data with number of variables p larger than the number of observations n occurs more frequently, it calls for advanced classification techniques. In Chapter 2, I proposed a novel sparse LDA method which generalizes LDA through a regularized approach for the two-class classification problem. The proposed method can obtain an accurate classification accuracy with attractive computation, which is suitable for high dimensional data with p>n. In Chapter 3, I deal with the classification when the data complexity lies in the non-random missing responses in the training data set. Appropriate classification method needs to be developed accordingly. Specifically, I considered the "reject inference problem'' for the application of fraud detection for online business. For online business, to prevent fraud transactions, suspicious transactions are rejected with unknown fraud status, yielding a training data with selective missing response. A two-stage modeling approach using logistic regression is proposed to enhance the efficiency and accuracy of fraud detection. Besides the classification problem, data from designed experiments in scientific areas often have complex structures. Many experiments are conducted with multiple variance sources. To increase the accuracy of the statistical modeling, the model need to be able to accommodate more than one error terms. In Chapter 4, I propose a variance component mixed model for a nano material experiment data to address the between group, within group and within subject variance components into a single model. To adjust possible systematic error introduced during the experiment, adjustment terms can be added. Specifically a group adaptive forward and backward selection (GFoBa) procedure is designed to select the significant adjustment terms. / Ph. D.

Made available in DSpace on 2016-06-02T20:06:06Z (GMT). No. of bitstreams: 1 4300.pdf: 2695135 bytes, checksum: c7742258a75f77aa35ccb54abc3439fe (MD5) Previous issue date: 2012-03-16 / Financiadora de Estudos e Projetos / In credit scoring problems, the interest is to associate to an element who request some kind of credit, a probability of default. However, traditional models uses samples biased because the data obtained from the tenderers has only clients who won a approval of a request for previous credit. In order to reduce the bias sample of these models, we use strategies to extract information about individuals rejected to be able to infer a response, good or bad payer. This is what we call the reject inference. With the use of these strategies, we also use the bagging technique (bootstrap aggregating), which consist in generate models based in some bootstrap samples of the training data in order to get a new predictor, when these models is combined. In this work we will discuss about some of the combination methods in the literature, especially the method of combination by logistic regression, although little used but with interesting results.We'll also discuss some strategies relating to reject inference. Analyses are given through a simulation study, in data sets generated and real data sets of public domain. / Em problemas de credit scoring, o interesse é associar a um elemento solicitante de algum tipo de crédito, uma probabilidade de inadimplência. No entanto, os modelos tradicionais utilizam amostras viesadas, pois constam apenas de dados obtidos dos proponentes que conseguiram a aprovação de uma solicitação de crédito anterior. Com o intuito de reduzir o vício amostral desses modelos, utilizamos estratégias para extrair informações acerca dos indivíduos rejeitados para que nele seja inferida uma resposta do tipo bom/- mau pagador. Isto é o que chamamos de inferência dos rejeitados. Juntamente com o uso dessas estratégias utilizamos a técnica bagging (bootstrap aggregating ), que é baseada na construção de diversos modelos a partir de réplicas bootstrap dos dados de treinamento, de modo que, quando combinados, gera um novo preditor. Nesse trabalho discutiremos sobre alguns dos métodos de combinação presentes na literatura, em especial o método de combinação via regressão logística, que é ainda pouco utilizado, mas com resultados interessantes. Discutiremos também as principais estratégias referentes à inferência dos rejeitados. As análises se dão por meio de um estudo simulação, em conjuntos de dados gerados e em conjuntos de dados reais de domínio público.

Abstracts in English, Zulu and Southern Sotho / The advent of Basel II Capital Accord has revolutionised credit risk measurement (CRM) to the extent that the once “perceived riskier bank assets” are now accommodated for lending. The Small and Medium Enterprise (SME) sector has been traditionally perceived as a riskier and unprofitable asset for lending activity by Commercial Banks, in particular. But empirical studies on the implementation of the Basel II internal-ratings-based (IRB) framework have demonstrated that SME credit risk is measurable. Banks are still finding it difficult to forecast SME loan default and to provide credit to the sector that meet Basel’s capital requirements. The thesis proposes to construct an empirical credit risk measurement (CRM) model, specifically for SMEs, to ameliorate the adverse effects of SME credit inaccessibility due to high information asymmetry between financial institutions (FI) and SMEs in Zimbabwe. A well-performing and accurate CRM helps FIs to control their risk exposure through selective granting of credit based on a thorough statistical analysis of historical customer data. This thesis develops a CRM model, built on a statistically random sample, known-good-bad (KGB) sample, which is a better representation of the through-the-door (TTD) population of SME loan applicants. The KGB sample incorporates both accepted and rejected applications, through reject inference (RI). A model-based bound and collapse (BC) reject inference methodology was empirically used to correct selectivity bias inherent in CRM domain. The results have shown great improvement in the classification power and aggregate supply of credit supply to the SME portfolio of the case-studied bank, as evidenced by substantial decrease of bad rates across models developed; from the preliminary model to final model designed for the case-studied bank. The final model was validated using both bad rate, confusion matrix metrics and Area under Receiver Operating Characteristic (AUROC) curve to assess the classification power of the model within-sample and out-of-sample. The AUROC for the final model (weak model) was found to be 0.9782 whilst bad rate was found to be 14.69%. There was 28.76% improvement in the bad rate in the final model in comparison with the current CRM model being used by the case-studied bank. / Isivumelwano seBasel II Capital Accord sesishintshe indlela yokulinganisa ubungozi bokunikezana ngesikweletu credit risk measurement (CRM) kwaze kwafika ezingeni lapho izimpahla ezazithathwa njengamagugu anobungozi “riskier bank assets” sezimukelwa njengesibambiso sokuboleka imali. Umkhakha wezamaBhizinisi Amancane naSafufusayo, phecelezi, Small and Medium Enterprise (SME) kudala uqondakala njengomkhakha onobungozi obukhulu futhi njengomkhakha ongangenisi inzuzo, ikakhulu njengesibambiso sokubolekwa imali ngamabhange ahwebayo. Kodwa izifundo zocwaningo ezimayelana nokusetshenziswa nokusetshenziswa kwesakhiwo iBasel II internal-ratings-based (IRB) sezikhombisile ukuthi ubungozi bokunikeza isikweletu kumabhizinisi amancane nasafufusayo (SME) sebuyalinganiseka. Yize kunjalo, amabhange asathola ukuthi kusenzima ukubona ngaphambili inkinga yokungabhadeleki kahle kwezikweletu kanye nokunikeza isikweletu imikhakha enemigomo edingekayo yezimali kaBasel. Lolu cwaningo beluphakamisa ukwakha uhlelo imodeli ephathekayo yokulinganisa izinga lobungozi bokubolekisa ngemali (CRM) kwihlelo lokuxhasa ngezimali ama-SME, okuyihlelo elilawulwa yiziko lezimali ezweni laseZimbabwe. Imodeli ye-CRM esebenza kahle futhi eshaya khona inceda amaziko ezimali ukugwema ubungozi bokunikezana ngezikweletu ngokusebenzisa uhlelo lokunikeza isikweletu ababoleki abakhethekile, lokhu kususelwa ohlelweni oluhlaziya amanani edatha engumlando wekhasimende. Imodeli ye-CRM ephakanyisiwe yaqala yakhiwa ngohlelo lwamanani, phecelezi istatistically random sample, okuluphawu olungcono olumele uhlelo lwe through-the-door (TTD) population lokukhetha abafakizicelo zokubolekwa imali bama SME, kanti lokhu kuxuba zona zombili izicelo eziphumelele kanye nezingaphumelelanga. Indlela yokukhetha abafakizicelo, phecelezi model-based bound-and-collapse (BC) reject-inference methodology isetshenzisiwe ukulungisa indlela yokukhetha ngokukhetha ngendlela yokucwasa kwisizinda seCRM. Imiphumela iye yakhombisa intuthuko enkulu mayelana namandla okwehlukanisa kanye nokunikezwa kwezikweletu kuma SME okungamamabhange enziwe ucwaningo lotho., njengoba lokhu kufakazelwa ukuncipha okukhulu kwe-bad rate kuwo wonke amamodeli athuthukisiwe. Imodeli yokuqala kanye neyokugcina zazidizayinelwe ibhange. Imodeli yokugcina yaqinisekiswa ngokusebenzisa zombili indlela isikweletu esingagculisi kanye negrafu ye-Area under Receiver Operating Characteristic (AUROC) ukulinganisa ukwehlukaniswa kwamandla emodeli engaphakathi kwesampuli nangaphandle kwesampuli. Uhlelo lwe-AUROC lwemodeli yokugcina (weak model) lwatholakala ukuthi luyi 0.9782, kanti ibad rate yatholakala ukuthi yenza i-14.69%. Kwaba khona ukuthuthuka nge-28.76% kwi-bad rate kwimodeli yokugcina uma iqhathaniswa nemodeli yamanje iCRM model ukuba isetshenziswe yibhange elithile. / Basel II Capital Accord e fetotse tekanyo ya kotsi ya mokitlane (credit risk measurement (CRM)) hoo “thepa e kotsi ya dibanka” ka moo e neng e bonwa ka teng, e seng e fuwa sebaka dikadimong. Lekala la Dikgwebo tse Nyane le tse Mahareng (SME) le bonwa ka tlwaelo jwalo ka lekala le kotsi e hodimo le senang ditswala bakeng sa ditshebetso tsa dikadimo haholo ke dibanka tsa kgwebo. Empa dipatlisiso tse thehilweng hodima se bonweng kapa se etsahetseng tsa tshebetso ya moralo wa Basel II internal-ratings-based (IRB) di supile hore kotsi ya mokitlane ya SME e kgona ho lekanngwa. Leha ho le jwalo, dibanka di ntse di thatafallwa ke ho bonelapele palo ya ditlholeho tsa ho lefa tsa diSME le ho fana ka mokitla lekaleng leo le kgotsofatsang ditlhoko tsa Basel tsa ditjhelete. Phuputso ena e ne sisinya ho etsa tekanyo ya se bonwang ho mmotlolo wa kotsi ya mokitlane (CRM) tshebetsong ya phano ya tjhelete ya diSME e etswang ke setsi sa ditjhelete (FI) ho la Zimbabwe. Mmotlolo o sebetsang hantle hape o fanang ka dipalo tse nepahetseng o dusa diFI hore di laole pepeso ya tsona ho kotsi ka phano e kgethang ya mokitlane, e thehilweng hodima manollo ya dipalopalo ya dintlha tsa histori ya bareki. Mmotlolo o sisingwang wa CRM o hlahisitswe ho tswa ho sampole e sa hlophiswang, e leng pontsho e betere ya setjhaba se ikenelang le monyako (TTD) ya batho bao e kang bakadimi ba tjhelete ho diSME, hobane e kenyelletsa bakopi ba amohetsweng le ba hannweng. Mokgwatshebetso wa bound-and-collapse (BC) reject-inference o kentswe tshebetsong ho nepahatsa tshekamelo ya kgetho e leng teng ho lekala la CRM. Diphetho tsena di bontshitse ntlafalo e kgolo ho matla a tlhophiso le palohare ya phano ya mokitlane ho diSME tsa banka eo ho ithutilweng ka yona, jwalo ka ha ho pakilwe ke ho phokotseho ya direite tse mpe ho pharalla le dimmotlolo tse hlahisitsweng. Mmotlolo wa ho qala le wa ho qetela e ile ya ralwa bakeng sa banka. Mmotlolo wa ho qetela o ile wa netefatswa ka tshebediso ya bobedi reite e mpe le mothinya wa Area under Receiver Operating Characteristic (AUROC) ho lekanya matla a kenyo mekgahlelong a mmotlolo kahare ho sampole le kantle ho yona. AUROC bakeng sa mmotlo wa ho qetela (mmotlolo o fokotseng) e fumanwe e le 0.9782, ha reite e mpe e fumanwe e le 14.69%. Ho bile le ntlafalo ya 28.76% ho reite e mpe bakeng sa mmotlolo wa ho qetela ha ho bapiswa le mmotlolo wa CRM ha o sebediswa bankeng yona eo. / Graduate School of Business Leadership / D.B.L.