Enhancing Credit Risk Analysis in Financial CRMs using Retrieval-Augmented Generation and Predictive AI

Abstract

Credit risk evaluation has always been a crucial part of the decision-making process in finance. However, researchers face significant challenges trying to apply statistical models that are not very malleable when it comes to reflecting the ever-changing behavior among borrowers or the changes in the economic environment. A recent development in machine learning, which can be categorized under Artificial Intelligence, is the Retrieval-Augmented Generation (RAG). This paper proposes integrating RAG with predictive AI models in credit risk analysis for financial CRM systems. Thus, RAG systems enhance structured financial datasets since the derived variables are real-time, borrower-specific and macroeconomic, enhancing the accuracy and responsiveness of financial dataset predictions. Therefore, as shown from the evaluations of the ChinaZJB SME dataset as well as other UCI benchmark datasets, the integration of RAG improves the roots of traditional prediction models, including Random Forest, Neural Networks, Support Vector Machines and Logistic Regression in terms of accuracy, recall and AUC-ROC. Further, the timely incorporation of new datasets reduces class imbalance and enhances minor class prediction beyond 20%. It describes the system, which includes the main elements, methods of data preparation, methods of training, methods of building and selecting the model and data privacy issues. These studies establish that integrating retrieval systems with AI shifts the landscape of financial CRM technology to the next level of solutions that can provide better customer experiences and agile and robust credit risk predictions. The future development can be the progress in developing an efficient and optimized retrieval system, integrating risk modelling for personal clients and using the system in other financial decision-making.

Keywords: Credit Risk Analysis, Retrieval-Augmented Generation (RAG), Predictive AI, Financial CRM Systems, Machine Learning, Risk Modeling.

1. Introduction

Credit risk management has become a key strategy that applies to the lending process to help manage and minimise the probability of default on a loan to maintain financial stability. Consequently, conventional risk measurement techniques primarily skewed toward historical records and other financial ratios are inadequate in an unstable and globalized world economy. With increasing competition, financial institutions are now more focused on evaluating credit risk and for this purpose more effective and efficient, accurate and interpretable models are also required^1-3. Using dispensation, client data storage and management, Fruit has developed from simple data storage and client database retainers into comprehensive customer-contact response, portfolio monitoring and decision-support tools. However, advanced credit risk analysis in CRMs has been limited, especially due to inadequate scoring models and inflexible data processing mechanisms. In order to tackle this deficit, there is a significant potential for amending the existing financial CRMs by implementing innovative, adaptive risk evaluation mechanisms that utilize voluminous real-time data environments.

Retrieval-augmented generation (RAG) can be considered the artificial intelligence breakthrough that incorporates large language models with an information retrieval mechanism. By integrating such data from outside the model, RAG offers a more up-to-date and broader base for credit risk assessment. With the right AI models for borrower behavior, macroeconomic signals and portfolio characteristics integrated to Risk risk-adjusted growth, the effectiveness and dynamism of risk management in a financial CRM can be elevated by RAG. This work unveils a new approach to combining Retrieval-Augmented Generation and predictive methods into financial CRM systems for intelligent credit risk analysis. The proposed architecture simultaneously boosts the prediction performance, increases the interpretability of the results and optimizes the computations. This paper outlines the key aspects of the system design, such as data source identification, data retrieval approach, model incorporation and system efficiency assessment. Concerning research findings attained from real-world financial datasets, the presented approach can outcompete previous risk models, permitting financial organizations to yield more accurate, quicker and more reliable credit decisions within a rising-native financial world.

2. Related Work

2.1. Traditional credit risk models

Traditional credit risk models have been organizations’ primary framework in evaluating default risk, loss frequencies and regulatory capital charges. Models like Credit Metrics™ used by J.P. Morgan or Credit Risk +™ of Credit Suisse companies follow the default mode or the mark-to-market approach to determine the behavior of credit portfolios. Such systems primarily refer to historical data for borrowers, internal credit scores and non-changing financial ratios for risk measurement. A characteristic that most of these model’s share is that they are unconscious models of credit risk where comparisons between borrowers’ characteristics and their defaults are relatively static or invariant to credit cycles^4-7. However, it becomes ineffective, especially in complex or changing economic scenarios of the Italian economy, which are dynamic. They cannot take and respond to real-time cues that the economy provides, for instance, changes in interest rates, geopolitical risk or a particular sector’s decline. Reliance on past information limits them in evaluating non-conventional borrowers or fresh financial risks. Especially in economic distress, these constraints will result in under or overestimating risks, portfolio vulnerability and regulatory compliance.

2.2. Use of AI and machine learning in risk assessment

AI and ML have revolutionized credit risk analysis by improving ways to inspect data and find unknown relations within big data space. Tools like Random Forest, Support Vector Machines (SVMs) and different types of Neural Networks also help risk analysts to integrate all sorts of structured & unstructured hard & soft data records like Transaction history, Geo mapping data and Behavioral Patterns from Social media posts. Random Forests are much more likely to detect more complex borrower risk factors than a single decision tree. Deep Neural Networks are very proficient in modelling complex variable interactions. However, the implementation of these models has been subject to criticism due to the interpretability of the models. AI algorithms operate based on unexplained models, making it challenging for financial institutions to explain credit decisions to regulators and auditors. To this end, methods such as Shapley values and LIME (Local Interpretable Model-Agnostic Explanations) have been created to provide much more interpretability without much reduction in accuracy. Nevertheless, critics of Artificial Intelligence have raised issues of data privacy, bias and fairness in the model; however, they have proven their competence in bringing down default rates, improving credit scoring and credit risk assessment, as well as extending credit accessibility to under-banked and unbanked populations hence revolutionizing the risk assessment market.

2.3. Retrieval-augmented generation (RAG) in financial applications

Retrieval-augmented generation (RAG) is a crucial development, particularly FOR artificial intelligence applications entailing action-based creativity, which dynamically correlates with existing information. RAG architectures enable models to retrieve the latest information from databases and/or the web to provide accurate and timely information in responses to queries processed. The large language models (LLMs) introduced into the RAG architectures provide an ability to bring external data for the current and accurate output. In a credit risk analysis, RAG allows an organization’s CRM systems to integrate up-to-the-minute financial conditions, changes in regulations and borrower information into its risk assessment processes. For instance, when using a RAG-enhanced CRM, the ability to obtain up-to-date macroeconomic factors such as unemployment rates, inflation rates or policy shifts together with borrower-specific changes such as revised income or spending are used to fine-tune the risk measurements even further. This addresses the knowledge cutoff major drawback that has been realized in most AI models and also ensures that there are minimal hallucinations because the answer is obtained from an authoritative source. Automated compliance reporting, anti-money laundering (AML) screening and personalized investment advisory are also some of the applications many financial institutions use RAG for, apart from credit scoring. Since RAG offers tangible and transparent results supported by proof, the system is a pioneering tool in the metamorphosis of smart, compliance-conscious and adaptive financial CRM solutions.

3. Proposed Methodology

3.1. System architecture overview

The architecture of the proposed approach that enriches credit risk analysis in financial CRM systems comprises real-time data ingestion, Retrieval-Augmented Generation (RAG) for dynamic contextual retrieval and advanced Predictive AI models for risk classification. The overall system is presented in Figure 1 below and is designed to prevent such drawbacks of a static model by regularly incorporating external financial indicators and customer behavior signals into the risk assessment^8-12. This can be achieved by designing the application modularly, where data flows from external sources, through individual CRM system components, into RAGs and then to the predictive analytic engines. At the base, the system takes multiple raw, non-structured inputs in financial transactions, raw credit bureau feeds, communication records and other external signals, such as market updates or fraud trends. Over time, it is kept in the CRM Data Warehouse with in-house information such as customer database risk evaluation. The results of each query are query embeddings that express a certain risk profile of the customer, which allows for information search and retrieval by the RAG module within a particular context.

The Retrieval-Augmented Generation (RAG) is designed to act as the final link between static CRM data and real external knowledge. As an information search technology, it processes analyst queries, searches for relevant documents and provides rich contextual data. This output is then entered into the Predictive AI Risk Analysis engine, which means that credit risk scores are not only based on static attributes of a borrower but, in addition to that, are supplied with real financial and behavioral data. For feature engineering, risk scoring and feature explanation, the predictive engine uses a combination of XG Boost, deep neural nets (DNN) and decision trees. In order to ensure that the assembling remains stable and the model remains relevant to the selected system, there is the Model Management and Feedback Loop responsible for evaluating the performance and checking for model shifts. Whenever there is a change in dataset characteristics or model performance, the regulation of retraining processes, update of the deployed models and registration happens systematically. Records are updated constantly in an interactive CRM dashboard, which enables the risk taker to either ratify or dispute the risk scores assigned by the application, thus eliminating the human-in-the-loop dilemma.

Last, the User Interaction and Monitoring module includes the real-time risk profile visualization, the risk score rationale and the heat-map interface. The analysts can look at specific customer data tables with selected fields, work with new risk alerts to follow and provide feedback that improves the algorithms used in the models (Figure 1). This synergy makes it possible to improve the accuracy and timeliness of credit risk analysis while simultaneously making it more transparent and easier to audit in a world that is ever more concerned with the explainability and fairness of the actions undertaken using Artificial Intelligence.

Figure 1: System Architecture for Enhancing Credit Risk Analysis in Financial CRMs using RAG and Predictive AI.

3.2. Data sources and preprocessing for CRM systems

The modern approach to credit risk analysis in connection with the use of modern CRM systems involves using many various but high-quality sources of information. The proposed framework receives data from conventional finance sources and new advanced sources. The external data include the financial transaction databases, credit bureau APIs, customers’ interaction logs and third-party risk signals. Each provides value for real-world spending patterning, scoring normalized for risk benchmarking and customers’ raw interacting patterns and sentiment. However, preprocessing is always required before the datasets can be used in analytical models. Data preprocessing techniques are practiced to deal with problems related to missing data, inconsistent data format and errors in entries. Normalization and feature scaling normalize or scale the numbers involved in order to bring all attributes from different sources to the same order. Thus, for the case of unstructured data like customer logs/risk alerts, features are derived from them using natural language processing (NLP). Furthermore, the entity resolution methods help to connect the customer data from various sources and enrich the consolidated customer view in the CRM Data Warehouse. This is very important to enhance the quality of the data to allow subsequent processes, such as the RAG module and the Predictive AI Engines, to run without interruption.

3.3. Retrieval-augmented generation (RAG) for contextual credit analysis

The RAG module is critical in changing the CRM data from passive to active and contextual credit risk information. In contrast to other dialogue modular systems, RAG does not contain pre-stored information required for the credit assessment. Still, it dynamically builds its context from real-time documents important for the current assessment task. When the Query Encoder component receives a query that can be consistent or created by the risk analyst based on some observed situation, the component takes the representation of an analyst’s query and the context of the customer. This vector is then given to the Document Retriever, which retrieves the updated risk documents relevant to the identified sources from both internal and external sources that have been indexed. The documents are then compiled into information packages by the Context Generator, including recent trends in finance behavior, regulation changes and the macroeconomic environment. In this way, RAG excludes the impact of such key issues as high dependence on outdated information and high probabilities of appearing hallucinations, which are inherent to the work of traditional language models. This real-time contextualization makes it easier to get updates about the borrower's current risk environment, making credit risk scores timelier and more accurate for financial decision-making.

3.4. Predictive AI models for risk scoring

The Predictive AI Risk Analysis engine performs the classic quantitative risk rating if the input is contextualized. In order to start with, there exists the Feature Engineering Engine that works by converting the big contextual data into vectors suitable for the efficiency of machine learning models. The interactions of features, which include feature crossing and embedding layers as well as the use of feature extraction tools such as principal component analysis (PCA). The core classifier is an ensemble of XG Boost, deep neural networks (DNNs) and decision trees for improved accuracy and stability of the model. After the credit risk score has been computed, the Explainability Engine adds brief explanations to its output for human understanding. The explanations on risk assessment are based on methods such as Shapley value decomposition and attention heatmaps that focus on the important features and events that led to the final decision. It also assures compliance with the legal aspects, thus increasing the level of trust between the lenders and borrowers and, at the same time enables the analyst to arrive at quicker decisions about the loans to offer. The proposed model architecture of the hybrid allows the avoidance of overfitting, the ability to detect new risk patterns and integration with continuous learning.

3.5. Integration strategy into financial CRMs

Seamless integration into existing financial CRM systems is critical for maximizing the utility of the proposed framework. This way, the architecture is designed for modular deployment where Financial Institutions can implement the framework and components in stages depending on their technological advancement and business needs. The main role of the CRM Data Warehouse is to provide mediation and integration, receive data from the external data ingestion layers and the RAG module and work with the predictive analytics engine and users’ dashboards through APIs only. These are procedures such as encryption, access control and logging to ensure that customer’s information is secured throughout the data flow process. In order to provide a good user experience, the Risk Alerts Module and the UI for the CRM Dashboard contain alerts, clickable heat maps and search functions to help the user find a specific customer. Notably, the Analyst Feedback Collector is a part of the CRM, which helps update the model, achieving the human-in-the-loop model and improving the level of risk scores. Using the Model Management and Feedback Loop means that models are updated as often as needed based on market changes. Using xAI means that outcomes obey the Fair Credit Reporting Act (FCRA) and Basel III benchmarks. Therefore, the integration aims to enhance incumbent CRM solutions whilst avoiding service interruption to facilitate the building of a smarter, more flexible and less vulnerable credit risk management framework.

4. Implementation Details

4.1. Technology stack and tools used

The technologies used in implementing the proposed credit risk analysis system are based on the technological stack that will allow for scalability and performance of the analysis, as well as how easily the technology can be integrated with the overall system^13-16. For data storing and processing, cloud relational databases like Amazon RDS and Azure SQL are used and data repositories such as Amazon S3 are used for dealing with the large amount of semi-structured transaction logs and interaction data. Data preprocessing is conducted in Python with the help of Pandas, NumPy and spaCy for manipulating both structured and unstructured data. The Retrieval-Augmented Generation (RAG) module is developed using Hugging Face’s Transformers and FAISS (Facebook AI Similarity Search) to index documents and passages. For the predictive AI models, suitable frameworks like XG Boost, TensorFlow, Py Torch, et cetera are used to develop and deploy the machine learning models to handle high dimensional feature space. The tools used for developing the interactive CRM interface and dashboards include React.js, While the CRM platforms used are force.com (Salesforce) and Microsoft dynamics via RESTful APIs. Kubernetes is used to manage the containers of the services to easily deploy and scale them, while tools such as ML flow are used to track the models throughout the development process (Table 1).

Table 1: Technology Stack Used for Enhancing Credit Risk Analysis in Financial CRMs

4.2. Model training and fine-tuning processes

The processes of model training and model fine-tuning are extremely important in order to enhance the ability of the system to predict accurately as well as to be robust over the long run. First, the baseline models are trained on historical credit performance data in which default, delinquencies, repayment history and other aspects are labelled. Thus, for the Retrieval-Augmented Generation module, large language models are trained on the financial and credit report, regulatory bulletins and credit analyst notes to attune the model to process financial jargon and focus on that particular type of language. Fine-tuning, on the other hand, works under transfer learning methods where the transformer models are trained on some other specific datasets with small learning rates during training to avoid the issue of catastrophic forgetting. Techniques such as grid search and Bayesian optimization are used to finalize model structures so that the chosen performance measures undergo optimization: AUC-ROC and the balance between precision and recall.

Furthermore, to cater for the problem of model drift that arises during the deployment process, a continuous training process is set up. This pipeline takes newly acquired risk events and feedback from the analysts to improve and update the models in the pipeline without needing to retrain the model from scratch. This technique enhances adaptability without interrupting operations.

Figure 2: AI Model Training and Fine-tuning Workflow.

4.3. Handling data privacy and compliance

Given the sensitive nature of financial and customer data, robust strategies for handling data privacy and compliance are integral to the system’s architecture. Customer-identifiable information (CII) is also preserved through the Secure Sockets Layer or Transport Layer Security protocol at both storage and transit with the use of protection schemes like Advanced Encryption Standard ciphering or AES-256 and Transport Layer Security stand TLS 1.3. RBAC is implemented at different levels within the system whereby only authorized people are allowed access or make changes to the sensitive data^17-20. As data is removed in tasks where possible, data anonymization and tokenization are applied where there is a need to replace specific data during the training phases. In order to ensure GDPR, CCPA and FCRA compliance, settings of audit trails and consents are automated. Audits of data access and penetration testing, as well as employing the services of third-party security assessors, are performed to reduce these risks ahead of time. The robust explainability architectures developed into the appropriately selected AI solutions can also support the reporting requirements to the regulators and enable reporting for the ruling for adverse consumer credit decisions.

5. Experimental Setup and Results

This section describes the experimental setting that will be employed to assess the performance of Retrieval-Augmented Generation (RAG) in credit risk analysis against traditional machine learning methods. It is detailed in the experiments that predictive performances, insensitivity to imbalanced classes and abilities for dynamic updating of models based on borrower information have been considered.

5.1. Dataset description

In order to enhance the validity of the study, this research employed two major data sets, The ChinaZJB SME dataset and several credit datasets from UCI. The ChinaZJB SME dataset stems from the annual loan ledger of a Chinese commercial bank and features the information of 1,329 SMEs, out of which 108 are defaulting and 1,221 are non-defaulting. The traditional analysis involves reports on financial ratios, additional loan details, credit rating and non-financial parameters. First of all, serious problems of class imbalance are notable as, on average, there is one default for 11 non-defaults. For assessing robustness across different environments, the work used UCI benchmark datasets. These comprise the Polish bankruptcy data (versions 1 to 3), the Australian credit approval data and the Taiwan credit card default data (Table 2).

Table 2: Summary of Datasets Used for Model Evaluation.

5.2. Evaluation metrics

In order to provide a holistic assessment of model performance, multiple evaluation metrics were utilized. Accuracy determined the number of correct classifications concerning the total number of cases. At the same time, precision assessed the capability of minimizing false positives by looking at the number of true positives against the total predictions made. Regarding recall, the emphasis was on the ability to detect a reasonable number of defaulters; it involved determining the percentage of actual defaulters that were accurately captured. The F1 Score, a combination of precision and recall indicators, was particularly used to analyze the results for markers for the imbalanced datasets. The AUC-ROC model was used to check the model's capacity in classifying between a defaulter and a non-defaulter using different threshold levels.

5.3. Baseline comparisons

The following compares the traditional machine learning models and the developed RAG-enhanced credit risk system. For training information only that was in structured data format, we had Random Forests, Neural Networks, Support Vector Machines (SVMs) and Logistic Regression models (Figure 3). At the same time, RAG systems used the data in structure formats plus dynamic data from unstructured databases (Table 3).

Table 3: Performance Comparison of Traditional Machine Learning Models.

Figure 3: Graphical Representation of Performance Comparison of Traditional Machine Learning Models.

5.4. Performance analysis of RAG vs traditional models

The introduction of Retrieval-Augmented Generation into the credit risk assessment pipeline yielded substantial improvements across all major evaluation metrics. Concerning the ChinaZJB SME classification task, the RAG-enhanced systems could provide results of around 95 % accuracy, which was much higher than the baseline Random Forest. This improvement mainly came from behavioral data of borrowers, transaction stability and sentiments from social media, which gave a real-time indication of the health of the borrowers.

RAG models achieved a better AUC-ROC of 0.96 than the Random Forests, with an AUC-ROC of 0.94 only. Indeed, this level of discriminative power is explained by the capacity of RAG to integrate real-time macroeconomic factors (such as employment statistics or sector-specific information) when it is evaluated. Further, RAG systems kept a precision of 0.93 and a recall of 0.90. Thus, false positives were very low while the right clients at risk were identified. Significantly, RAG was able to claim that their capacity to process external big data contributed greatly to the issue of class imbalance. These rates show the effectiveness of RAG-enhanced systems in contrast with the previous models and these had an increase of approximately 22% of the minority class (default prediction), which can help in the early identification of high-risk customers. Thus, enhancing the adaptive context-awareness of credit risk prediction requires the integration of the retrieval mechanisms with the machine learning model for more efficient credit risk prediction.

6. Discussion

The corresponding analysis of the results obtained from the experiment showed that using the Retrieval-Augmented Generation (RAG) in financial CRM systems significantly improves credit risk assessment. RAG makes models have situational awareness, as they can update their behavioral data of a specific borrower and macroeconomic indicators in real-time, which is not characteristic of traditional AI. This real-time contextualization also enhances the marksmanship and recall, as well as the identification of early warning signs of borrower’s behavior, which will help to promote the effectiveness of risk management. Especially for imbalanced data sets where defaults are few but important for detection, like our ChinaZJB, the RAG-augmented models have been found to have 22% higher Minority Class Recall, which can revolutionize how institutions could contain credit risk.

RAG-enhanced systems lead to enhanced performance levels; however, it brings new challenges. Managing external data acquisition in real-time processing truly calls for a well-organized, well-equipped framework, better and stringent validation and verification principles and considerable and sophisticated query optimization measures to achieve both reliability and control of latency. However, privacy, compliance (i.g.., GDPR, CCPA) and ethical usage of unstructured data will always be critical for maintaining regulators’ approval and clients’ trust. Interpretability mechanisms such as explainability engines are even more important when the model’s features or predictions depend on information that can change in a short time, like social media updates. Financial institutions, hence, have to consider the two elements of superior and optimum prediction as they also foster openness, responsibility and regulation of artificial intelligence-based decisions. Transitioning from RAG into predictive credit risk analysis is a great leap towards the next generation of financial CRM systems. This not only helps firms predict defaults more accurately but also makes them improve constantly in adjusting to the economic conditions that constantly change. Future work in this regard might consider a refinement of the retrieval approach, the application of reinforcement learning for better risk management and the use of federated learning for improved data privacy during the usage of downstream external data sources.

7. Future Work

The future advancements in the retrieval systems will enhance the efficiency, precision and relevance of the information retrieved during credit risk assessments. Many current approaches based on RAG models have issues such as high retrieval latency and noise when it comes to unstructured information. This can be achieved with the help of using such techniques as dense passage retrieval, knowledge graphs and multi-hop retrieval chains of the sources. Moreover, the inclusion of some kind of reinforcement learning-based feedback option can be introduced that would help make the learning process in the given retrieval systems dynamic by revealing which sources are most trustworthy and therefore more contextually relevant and helpful, for different types of borrowers by improving the precision and speed of risk predictions.

Personalized credit risk modelling is another area of growth in the field in the future. Standard credit scoring procedures provide little distinction between borrowers in the same risk profile or age. However, the next-generation systems are ideally positioned to create even truly accurate risk scores using RAG combined with much more detailed borrower characterization targeting not only transactional behavior, real-time social media phenomena and even borrowers’ economics. It would facilitate accurate differentiation of credit products, customized setting of the interest rate structures and the ability to monitor early warning signs of firms in distress for each customer, thus enhancing customer satisfaction and portfolio performance. However, the general idea of applying RAG and predictive AI in other aspects of financial decisions goes further than credit risk analysis. Some applications include investment advisory services, anti-money laundering (AML) tools, insurance and claims underwriting, particularly when it involves real-time retrieval. The following are some examples: active management of investment portfolios based on real-time data on geopolitics events and the use of the Internet to adjust insurance premiums based on lifestyle data. It is believed that as information retrieval systems continue to grow in terms of their capabilities in terms of scalability and context sensitivity, there will be significant growth in the use of these systems within the financial sphere to make more responsive or more data-driven decisions.

8. Conclusion

This study demonstrates the transformative potential of integrating Retrieval-Augmented Generation (RAG) with predictive AI models in the context of credit risk analysis for financial CRM systems. By Incorporating RAG, methods improve the modelling precision and recall and reduce sensitivity to class imbalance about more conventional machine learning solutions achieved by combining structured financial data and real-time external data retrieval. Hence, RAG’s capability of responding to borrower behaviors and macroeconomic changes makes it a critical tool in applying agile and smart credit risk management strategies.

RAG also incompletely solves important problems of descriptive AI, such as the knowledge cutoff problem and static risk profiling. Applying current values of the market changes in regulations and evaluating the borrowers ‘behavior signals lets the financial institutions adapt quickly to economic conditions. However, to adopt and implement RAG systems, revision of data governance, model explainability and regulating requirements are necessary to enhance trust. In conclusion, this paper presents RAG-enhanced predictive AI as a symbolic direction toward improving the financial decision system. Future research development will also focus on improving the retrieval technique, communicating risk assessment and bringing horizons to other segments of the financial industry. That way, such changes can make a difference for financial institutions to deliver higher accuracy, fairness and efficiency in risk assessment and client engagement.

9. References

1. Wang Z. Artificial Intelligence and Machine Learning in Credit Risk Assessment: Enhancing Accuracy and Ensuring Fairness. Open Journal of Social Sciences, 2024;12: 19-34.
2. Giudici P, Hadji-Misheva B, Spelta A. Network-based credit risk models. Quality Engineering, 2020;32: 199-211.
3. Mays E. Credit risk modelling: design and application. Global Professional Publishing, 1998.
4. Aziz S, Dowling M. Machine learning and AI for risk management. Springer International Publishing, 2019: 33-50.
5. Pedron CD, Picoto WN, Dhillon G, et al. Value-focused objectives for CRM system adoption. Industrial Management & Data Systems, 2016;116: 526-545.
6. Paltrinieri N, Comfort L, Reniers G. Learning about risk: Machine learning for risk assessment. Safety Science 2019;118: 475-486.
7. Hegde J, Rokseth B. Applications of machine learning methods for engineering risk assessment- A review. Safety Science, 2020;122: 104492.
8. Sharma A, Singh UK. Modelling of smart risk assessment approach for cloud computing environment using AI & supervised machine learning algorithms. Global Transitions Proceedings, 2022;3: 243-250.
9. Zhang B, Yang H, Zhou T, Ali Babar M, Liu XY. Enhancing financial sentiment analysis via retrieval augmented large language models, in Proceedings of the fourth ACM international conference on AI in finance, 2023: 349-356.
10. Pan J, Yang Q, Yang Y, et al. Cost-sensitive-data preprocessing for mining customer relationship management databases. IEEE Intelligent Systems 2007;22: 46-51.
11. Peppard J. Customer relationship management (CRM) in financial services. European Management Journal 2000;18: 312-327.
12. Siddiqi N. Intelligent credit scoring: Building and implementing better credit risk scorecards. John Wiley & Sons, 2017.
13. Ryals L, Payne A. Customer relationship management in financial services: towards information-enabled relationship marketing. Journal of strategic marketing, 2001;9: 3-27.
14. Zheng H, Shen L, Tang A, et al. Learn from model beyond fine-tuning: A survey, 2023.
15. Kamath CN, Bukhari SS, Dengel A. Comparative study between traditional machine learning and deep learning approaches for text classification. In Proceedings of the ACM Symposium on Document Engineering, 2018: 1-11.
16. Chauhan NK, Singh K. A review of conventional machine learning vs deep learning. In 2018 International Conference on Computing, power and communication technologies (GUCON), 2018: 347-352.
17. Delamaire L. Implementing a credit risk management system based on innovative scoring techniques (Doctoral dissertation, University of Birmingham), 2012.
18. Siddiqi N. Credit risk scorecards: developing and implementing intelligent credit scoring. John Wiley & Sons, 2012;3.
19. Wu TC, Hsu MF. Credit risk assessment and decision-making by a fusion approach. Knowledge-Based Systems, 2012;35: 102-110.
20. Doumpos M, Lemonakis C, Niklis D, et al. Analytical techniques in the assessment of credit risk. EURO advanced tutorials on operational research, 2019.