Kamran Asadov
Azerbaijan State Oil and Industry University
April 2026

In the banking sector, the use of machine learning, such as in credit scoring, can accelerate processes and enhance the quality of decisions, but it also introduces some tensions regarding transparency and accountability of decisions. SHAP and LIME are not solutions to this problem as they are methods to interpret black-box models. They give interpretations that may be contradictory and that are not a true reflection of the decision making process of the system step by step. A more promising alternative might be EBMs that can be interpreted at the level of their architecture and are accurate enough. In the case of emerging markets like Azerbaijan, it is appropriate to consider machine learning systems in credit scoring from the start, with these issues in mind, so that they are ready for the future when regulations are being introduced around the world and slowly trickling into developing countries. In this paper, the authors explore why the post-hoc explanation methods do not meet these needs, and argue that inherently interpretable models are a better foundation for the adoption of ML in banking.

Keywords: machine learning, model selection, credit scoring, auditability, explainable AI, emerging markets.

In the context of emerging markets such as Azerbaijan, the transition to the use of machine learning in decision-making by financial institutions is still at an early stage. Even today, many banks continue to use manual decision making processes for loan approvals that are based on rule-based algorithms. Using data sources other than historical data can also be a concern with ethical issues. Such data, such as payment history, utility payments, e-commerce activity, and digital wallets, should be transparent in models, and the results should be explainable in simple terms (World Bank, 2024).

Azerbaijan is not an exception, as banks in developing countries are trying to implement machine learning in different areas, such as credit scoring, to automate decision-making (Karimova, 2024). However, the question that needs to be asked is how can a bank make a successful and responsible transition from manual to automated decision making?

There is no easy answer to this question. Among the models that are widely used, black-box models such as gradient boosting and neural networks are the most accurate. These are the least transparent models when it comes to decision making. To solve the problem of interpretability of black boxes, there are some methods that are grouped under the umbrella of XAI. The most popular of these are SHAP and LIME. The idea of these methods is to make the decision making of black-box models more comprehensible by providing explanations after the model has generated a result; this is a post-hoc approach. But many studies cast doubt on the reliability of such methods. The results of these methods have been found to be inconsistent across different runs in the same model, often do not reflect the model's internal processes and, overall, this approach fails to meet the requirements of regulatory frameworks like the EU AI Act and the GDPR.

This paper is unequivocal on this issue: relying on post-hoc XAI interpretation methods is not recommended for emerging market banks, and especially for banks in Azerbaijan, which are starting to use AI in credit scoring. A more flexible and reliable route is to employ models that are interpretable in the model itself. Specifically, these models are transparent in their design and are as accurate as black-box models, such as Explainable Boosting Machines (EBMs). In this paper, the author argues that it is better to take a more conservative and defensible approach and look to the future, than to have to deal with issues of interpretation and answerability later on for what has been done.

Increasingly, machine learning is playing a crucial role in financial institutions' decision-making processes, if not a critical one. This involves credit scoring, fraud detection, risk assessment and management. The research demonstrates that more complex models like gradient boosting and neural networks can make more accurate decisions, but come with a price: transparency and answerability.

According to the EU AI Act, credit scoring falls into the high-risk category of operations. For many years, the U.S. Federal Reserve has insisted on the careful testing of models that are put into practice. These regulations are not merely recommendations. In the meantime, the technical tools employed to explain the workings of black-box models have come under scrutiny in recent years, both because of their lack of transparency and because they do not meet the regulators' demand for answerability. In this respect, Azerbaijan is an interesting case: banks in the country are slowly starting to use machine learning in their decision-making processes.

Concurrently, the initial steps towards regulation provide a helpful backdrop for examining the issue of accuracy and answerability, and the quest for solutions in developing markets is one of the more pressing challenges in this field.

This paper aims to:

Is the post-hoc explanation enough for responsible use of ML in high-stakes banking decisions, and when should inherently interpretable models be used?

The rest of this paper is organized as follows. Section 2 summarizes the pertinent literature on four fronts: regulatory lag and model reliability gaps, the use of ML in banking and related compliance pressures, technical challenges of post-hoc XAI methods, and the regulatory landscape of algorithmic decision-making. The main results of the paper are summarized in Section 3, bringing together the evidence for and against the post-hoc methods, and for the alternative approach of using inherently interpretable models, in particular EBMs, as a more accountable method. The last part of Section 4 contains implications for banks in Azerbaijan and other emerging market settings, as well as suggestions for future research.

The dilemma of Regulation when applying and implementing AI in financial services (emerging markets in particular), has failed to catch up with the pace of adoption of such technologies. Further, as financial market in Azerbaijan is continuing to undergo structural transformations, it’s been failing to develop proper regulation for Algorithmic Decision-Making Systems (Huseynova & Balajayeva, 2025; also the authors underline the consumer and confidence aspect as well as risks, protection of personal data fails to follow the process of super-automation of lending. We can see the same phenomenon from a technological perspective, too).

As an example Karimova (2024) shows that although in Azerbaijan Credit Scoring process is conducted mainly by manual credit scoring models, Banks will step into smarter ML applications until there is a point in Karimova's study 'black-box' models get better performance than classical methods but have difficulty explaining its output “whether should the black box models be used for scoring on a default basis given low of quality of data available and capacity for self modeling to develop credit scoring policies.” The problem becomes critical when tools that generally relied on dealing this problem get undermined. In fact, SHAP and LIME as explanatory post- hoc tools proved unreliable in generating different interpretations among different state-of-the-art models at a given case-Pombra (2022)-“raise questions about their suitability for compliance workflows”. Problem becomes even complicated with alternate data sources; as stated by the world bank(2024), “the more an agency uses credit assessment based on alternate data sources (e.g. Telephone bills, utility bills, etc.) the less it may matter what criteria the agency uses to set decision parameters on the business it wishes to conduct with its customers.” In short, emerging economies, such as Azerbaijan, are often challenged to effectively harness the benefits of ML approaches for commercial reasons while simultaneously ensuring compliance and avoiding the use of technologies that could result in a range of violations.

Ayari et al. (2025) reviewed credit scoring models in 63 articles published from 2018 to 2024 and found that models of ensemble and hybrid type show better performance. While deep learning holds great promise, it is characterized by limitations regarding interpretability and transparency issues and there are emerging demands in the industry such asSHAP, and LIME due to the needs for the latter, with its main limitations being data privacy, interpretation and transparency, complexity of computations and mitigation of biases. Other studies highlight, in addition, the tensions between performance and transparency Jacobs (2024).

A comparison is made between interpretable machine learning (IML) techniques (neural network with ReLU as activation function, GAMI-Net and Explainable Boosting Machines (EBM)) with logistic regression, applied to predict Probability of Default in historical corporate loan data, which concludes that models more sophisticatedthan logisticregressiononly obtain a minor improvement in perform, especially on test samples, casting doubt on the value of discarding transparency in pursuit of a tiny performance improvement. Moreover, EBM provides satisfactory performance results in training set, however performs poorly compared to other methods on unknown datasets (test set). Bussmann et al. (2020) present a practically apply a post-hoc explainability method, XAI, on their predictive model for default probability based on an XGBoost classifier on data originated from the peer-to-peer lending sector.

Their predictive model shows accuracy superior to that of logistic regression while XAI provides an insight on the relative influence of each feature on the specific prediction. While this can be viewed as a work in support of the paradigm "black box+post-hoc explanation", further sections illustrate why explainability only partially satisfies the accountability requirements as per regulations. Rudin (2019) provides some of the main arguments against relying on post-hoc explanation. According to her, the assumption of performance-transparency trade-off is often false.

With sufficiently descriptive and well-preprocessed features, some IMLs can compete with black-box solutions. Discrepancies in accuracy between black-boxes and IMLs often come from data pre-processing or from the whole construction of the problem and the features, rather than model complexity alone. Otherwise, She points out that the explanatory models do not provide the true mechanism of the original model, instead, provide only a partial and thus incomplete reflection of the decision-making. The term “explanation” used for these methods might be deceptive: the output of such methods are not the explanation for real decision, but interpretation of a surrogate model that approximately describes the decision process.

This is not an adequate answer when real answerability is expected by regulators. This last assertion is backed up by the work of Bhatt et al. (2020), which investigate the real impact of deployment of XAI systems in around thirty organizations (including financial intuitions) in an array of domains; their results show that existing XAI systems mainly benefit the internal data science community.

A fundamental assumption that underpins the popular usage of post-hoc explanation tools like SHAP and LIME is that these tools provide reliable information on why a specific model made a given decision. This assumption has become highly problematic in the literature. Pombra (2022) defines the disagreement problem, where different tools that are applied to the same model to explain the same decision produce inconsistent results. This issue does not pose minor technical limitations.

In sensitive domains, such as credit scoring where a bank must account for the denial of a loan, contradictory explanations make it impossible to put forth a single valid account of the decision.

The difficulty is more than one tool differing from another. Freiesleben and Konig (2023) highlight that most XAI methodologies sample the model at points where it did not see any of the original data, thus meaning that an explanation given is an extrapolation not a true representation of what the model is actually doing. A single explanation technique could never work reliably even for this because a model would not necessarily extrapolate correctly to unobserved points. Research on what happens in real world applications of explainability in the financial sector by Bhatt et al.

(2020), which reviewed the usage of XAI tools across about thirty different organizations including banks and credit bureaus, revealed that the usage of LIME explanations in reality gives results that the authors did not expect, which deviated from human reasoning and this was shown to arise from sampling variance and sample size. Even more telling, the authors found that the main function of these explainability tools in financial organizations is as debugging aids for in-house ML engineers, not for end users or regulators, whom they are often presented as serving.

This discrepancy in usage reveals a conflict between what XAI tools are said to do and what they in fact do: in reality, they hinder rather than help accountability, for they offer the illusion that decision-making has been made transparent and controllable without there being actual accountability for the decisions and people’s lives. Rudin (2019) demonstrates how explanation-seeking practices in the financial sector even threaten already existing legal safeguards such as GDPR, which demand explanation, because it cannot be given meaningfully, but a model that approximates the decision-maker. Therefore, the problems surrounding post-hoc explainability in sensitive areas are design-related rather than the current practice of the explainability tools, and, as is described in the next paragraph, the approach of integrating interpretability within the model is likely a more viable strategy (Chen et al., 2018).

Algorithmic decision-making is a topic that has now gone from academia to law. On the world stage, the European Union’s AI Act is probably one of the most comprehensive approaches of its kind so far. GDPR is providing assistance in a supporting role, by allowing people to opt-out of algorithmic decisions viaArticles 21 and 22, and forcing data controllers to have to justify its compliance in its use of data. Unfortunately, to know what a satisfactory response will actually look like, there are more issues to consider.

The debate about transparency versus true accountability continues to rage on (Busuioc, 2021), as although transparency is necessary for accountability, it isn’t quite sufficient. With complex algorithmic decisions that have stakes that high, understanding how the model is built in its generality does not mean much to an individual whose data has been used. How an individual data inputs lead to an individual output must be understood for true accountability to exist. The distinction between transparent SHAP models, and what accountability in AIs must look like cannot be overstated.

A more advanced model for conceptualizing accountability comes from Almada et al. (2024). They define accountability in AIs with reference to four different goals – namely compliance, report, oversight, and enforcement. As these authors discuss, even though theEU AI Act’s emphasis is onex-ante compliance, the other components (especially meaningful oversight and enforcement) remain largely undefined and underspecified.

For a market like Azerbaijan this has particular relevance. In countries such asAzerbaijan, there may be no national law governing algorithm use. However there are still many concerns about, and approaches to managing this risk. In Azerbaijan, in a country where it may not be legally binding on banks to employ interpretable models, the risk to a firm choosing a black-box model, with post-hoc explanation, if regulations shift to include the same rules as those developed in theEUis considerable (Huseynova & Balajayeva, 2025).

Those post-hoc explanation challenges that we spoke about in the last section are not a fundamental inevitability of machine learning. It specifically has to do with using explanation tools with models never intended to be understandable. It is more logical to use a model that builds interpretability into the model architecture from the start, making a separate explanation tool redundant. Of the available options, the Explainable Boosting Machine model looks like the most exciting advancement for tabular data in finance (like credit scoring).

A history The explainable boosting machine model is descended from the family ofGeneralized Additive Models (GAMs), models in which the prediction of an observation can be computed by summing a set of independent contribution terms associated with individual input features via a ‘shape’ function (that has been ‘learned’ from the data). , 2012, 2013), Lou and co-authors demonstrated that it is possible to adapt the GAM framework with advanced machine learning-boosting algorithms to create models that are more accurate than the original GAMS but remain perfectly transparent-the contributions for each feature are exact and verifiable without using any ‘surrogate’ models. They called these “enhanced” versions ofGAMs GA2M’s (Generalized Additive Models with pairs interactions). Fortunately, this concept has been put into code, available in an open-source Python library (InterpretML’s“EBM”model, Nori et al.

2019), which data scientists working with financial regulation know can easily be implemented. Countering the myth The ‘accuracy vs interpretability trade-off’ has been an ongoing debate in machine learning: the perception has been that to maximize accuracy, one must rely on the inscrutable “black box” models. Rudin(2019) counters this argument, asserting that in the realm of well-structured, tabular data (which constitutes the majority of datasets used in banking and credit decisioning), the tradeoff between accuracy and interpretability is, in practice, a myth. (2022) further the point through the notion of theRashomon set: a concept based on the observation that, for any real-world task, there may be a very large set of diverse models with near-equal predictive performance.

If that’s the case, there is absolutely no rational reason to choose the non-transparent options. Jacobs (2024) demonstrated empirically in credit that there is relatively little in-sample advantage for black-box models over interpretable models that offsets the loss of auditability out of sample. The evidence supporting this isn't limited to theoretical arguments. (2020), for example, demonstrated that it is possible to achieve similar levels of interpretability as traditional credit scorecards using machine learning-all while capturing full model power in real-world credit scoring contexts.

For them, transparency and interpretability should be an asset rather than a burden. (2018) applied an interpretable additive model onFICOchallenge data to achieve performance indistinguishable from neural networks, but with perfectly auditable, globally coherent explanations. A recommendation for emerging markets A key message for banks in Azerbaijan and other developing countries is this: given the choice between a model built to explain and a model requiring a post-hoc explanation, it is no longer necessary to choose between accuracy and interpretability. It becomes a choice between comparable predictive accuracy with fully auditable explanations versus comparable predictive accuracy without them.

In developing environments with lower regulatory maturity, using an inherently transparent model (like an EBM) is simply the sensible, responsible, and smart way forward.

Based on the review of literature, three closely related findings answer the research question.

The first finding is that algorithmic explainability and accountability are fundamentally distinct. Busuioc (2021), Rudin (2019) and Lima et al. (2022) differentiate them; explainability is a measure of the ability to produce a description of a model's outputs, whereas accountability refers to having the capacity for a decision to be subject to an audit, challenge and legal attribution. Traditional post-hoc explainability techniques like LIME and SHAP, can satisfy the former, but not the latter.

As demonstrated by Pombra (2022) and Krishna et al. (2022), even the same post-hoc explanation method, varying with inputs or algorithm parameters, will result in contradictory outcomes for the same output, leading to potential for misdirection when providing evidence or for the same decision to be rationalized in entirely different ways. An instance where the bank might claim to have a post-hoc explanation for a credit rejection through the use of SHAP values would fail to achieve the standards of answerability required of regulators. Freiesleben and Konig (2023) further reinforce these observations, as their findings suggest that the nature of contradictions encountered by post-hoc methods arises not from their application but the inherently flawed logic underpinning their use.

When probing models outside of areas where their training data lie, post-hoc models are unable to reliably and predictably reveal true reasoning, a core need in achieving accountability.

The second finding is that the accuracy trade-off argument is, to a large extent, a fallacy. The prevalent assumption among financial sector professionals is that opaque models perform significantly better predictors than more understandable options, justifying their use. Evidence presented challenges this; Rudin (2019, 2022) asserts that given structured tabular data, accuracy discrepancies are, in most instances, non-existent and that, most of the time, a set of equally accurate models include those that are transparent and interpretable. This notion is empirically confirmed by Jacobs (2024) for credit scoring, as a number of models with high performance are interpretably efficient, and in the applications addressed by Bucker et al.

(2020), the work of Fazel et al. (2026), EBM's are observed to have similar performances to opaque models but with transparency guaranteed. Without any inherent differences in performance between alternatives, there is no justification for opting for a model that is not interpretable.

The third finding is specific to the context of the Azerbaijan market and its banks, which are characterized as historically using manual and rule-based scoring and which will be forced to increasingly adopt more complex models. Banks operating in Azerbaijan will face growing demand to use such tools; however, the Azerbaijani legislative and regulatory framework regarding data use and algorithms remains inadequate, as documented by Huseynova and Balajayeva (2025). Under these conditions, the utilization of opaque models with post-hoc explainability is a financial institution not merely making an innovation choice but also a regulatory decision, in an environment that increasingly emphasizes such compliance on an international level. Even in well-established banking systems, emerging markets utilizing complex alternative data sources will be strained in their capacity to produce effective justifications for ML-driven choices (World Bank, 2024).

With these findings, it is apparent that in markets like Azerbaijan, inherently explainable models like EBM are not simply an alternative choice but a sensible and ethical default option because they not only match performance but build answerability from conception.

This paper sought to evaluate whether post-hoc explanations are sufficient for accountable ML deployment for banking high-risk decisions and under which circumstances one would rather opt for inherently interpretable models. The conclusion is supported by a body of evidence coming from technical, empirical, and regulatory literatures, which is overwhelmingly that EBMs are insufficient, and that the best foundation for emerging market banks intending to adopt ML for credit scoring are inherently interpretable models.

The argument formulated in the present study hinges on three premises. First, the distinction between explanation and accountability is a substantive difference of kind and not a matter of degree. Output explanations of a ML model do not equate to being truly answerable for a decision impacting an individual's financial well-being, and emerging global regulation, including both the EU AI Act and GDPR, is increasingly designed around the latter not the former. Second, the technical case in favor of ML models based on black boxes is arguably overstated in the domain of credit scoring.

On structured tabular data, the improvement in predictive performance afforded by complex black-box models is negligible over well-specified interpretable options, and for every comparably performing, complex model, there's always an equivalent interpretable option that would fall into the relevant Rashomon set. Finally, the context of banks in emerging markets such as Azerbaijan is that without the existence of an advanced framework of AI Regulation, and by adopting models which can shape regulatory requirements and choices on a long term basis, banks are making themselves accountable to decisions based on poor standards which are taken by researchers in more mature markets in their own regulatory and societal context.

A few recommendations for banks, as well as for local and international standard setting bodies and research communities can be derived. Banks in Azerbaijan which consider using ML for credit scoring in the high-risk area of credit scoring need to be approached from a governance viewpoint, with appropriate model validation. EBM should be used preferentially unless specific business needs cannot be met by these approaches, and there's solid evidence proving the superiority in prediction accuracy when using models based on black boxes and local context when it comes to the specific application. Regulators and Standard Setters such as AZSTAND should try to define what 'transparency of models', rather than 'output explainability' entails for the local context, so that to establish minimum interpretability standards for high-risk applications, in line with the ones implemented globally.

Researchers in the local context or in a more general developing markets setting need to carry out empirical studies in local conditions which would contrast different EBM approaches on relevant data, which is a most pressing gap, to address and analyze further.

This paper is subject to some limitations. Firstly, this paper presents a theoretical review and does not test any model empirically. The models have been taken from the research conducted in other geographical regions and local features of Azerbaijani financial data might have different effects on the models' performance. Second, EBM performance data in terms of predictions for credit scoring also based on nonAzerbaijani data sets, which implies the limitation related to performance evaluation criterion as established for Azerbaijan’s local circumstances, for which a more granular methodology should be sought.

However, all the limitations discussed above highlight the agenda for future research. To overcome the second limitations discussed, the present research recommends an empirical study focusing on an application of EBMs on a locally available dataset from Azerbaijan’s financial institutions. Such research would further support the claims made in this paper on the effectiveness of EBMs in fulfilling the requirements related to accountability standards, especially in the developing countries’ financial sector.

Almada, M., Lagioia, F., & Sartor, G. (2024). Accountability in AI: A structured analysis. AI & Society, 39, 1871-1882. https://doi.org/10.1007/s00146-023-01638-9

Ayari, H., Guetari, R., & Kraiem, N. (2025). Machine learning powered financial credit scoring: A systematic literature review. Artificial Intelligence Review, 59(1). https://doi.org/10.1007/s10462-025-11416-2

Bhatt, U., Xiang, A., Sharma, S., Weller, A., Taly, A., Jia, Y., Ghosh, J., Puri, R., Moura, J. M. F., & Eckersley, P. (2020). Explainable machine learning in deployment. In Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency (pp. 648-657). ACM. https://doi.org/10.1145/3351095.3375624

Bucker, M., Szepannek, G., Gosiewska, A., & Biecek, P. (2020). Transparency, auditability and eXplainability of machine learning models in credit scoring. arXiv. https://arxiv.org/abs/2009.13384

Busuioc, M. (2021). Accountable artificial intelligence: Holding algorithms to account. Public Administration Review, 81(5), 825-836. https://doi.org/10.1111/puar.13293

Bussmann, N., Giudici, P., Marinelli, D., & Papenbrock, J. (2020). Explainable AI in fintech risk management. Frontiers in Artificial Intelligence, 3, Article 26. https://doi.org/10.3389/frai.2020.00026

Chen, C., Lin, K., Rudin, C., Shaposhnik, Y., Wang, S., & Wang, T. (2018). An interpretable model with globally consistent explanations for credit risk. arXiv. https://arxiv.org/abs/1811.12615

Fazel, R. E., Bakhtiary, A., & Bigdeli, S. A. (2026). Improving credit card fraud detection with an optimized explainable boosting machine. arXiv. https://arxiv.org/abs/2602.06955

Freiesleben, T., & Konig, G. (2023). Dear XAI community, we need to talk! Fundamental misconceptions in current XAI research. In xxAI: Beyond Explainable AI (pp. 39-63). Springer. https://doi.org/10.1007/978-3-031-04083-2_3

Huseynova, S., & Balajayeva, L. (2025). Digital banking regulation in Azerbaijan: Empirical insights into user trust, risk management, and data protection. Journal of Eastern European and Central Asian Research, 12(2), 80-97. https://doi.org/10.15549/jeecar.v12i2.2153

Jacobs, M. (2024). Benchmarking alternative interpretable machine learning models for corporate probability of default. Data Science in Finance and Economics, 4(1), 1-52. https://doi.org/10.3934/DSFE.2024001

Karimova, N. (2024). Application of AI in credit risk scoring for small business loans: A case study on how AI-based random forest model improves a Delphi model outcome in the case of Azerbaijani SMEs. arXiv. https://arxiv.org/pdf/2410.05330

Krishna, S., Han, T., Gu, A., Pombra, J., Jabbari, S., Wu, S., & Agarwal, S. (2022). The disagreement problem in explainable machine learning: A practitioner's perspective. arXiv. https://arxiv.org/abs/2202.01602

Lima, L., Camarinha-Matos, L. M., & Novais, P. (2022). The conflict between explainable and accountable decision-making algorithms. In Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency (pp. 1-13). ACM. https://doi.org/10.1145/3531146.3533243

Lou, Y., Caruana, R., & Gehrke, J. (2012). Intelligible models for classification and regression. In Proceedings of the 18th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (pp. 150-158). ACM. https://doi.org/10.1145/2339530.2339556

Lou, Y., Caruana, R., Gehrke, J., & Hooker, G. (2013). Accurate intelligible models with pairwise interactions. In Proceedings of the 19th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (pp. 623-631). ACM. https://doi.org/10.1145/2487575.2487579

Nalbantov, G., Birbil, S. I., & Brinkhuis, M. (2024). Enhancing ML models interpretability for credit scoring. arXiv. https://arxiv.org/abs/2509.11389

Nori, H., Jenkins, S., Koch, P., & Caruana, R. (2019). InterpretML: A unified framework for machine learning interpretability. arXiv. https://arxiv.org/abs/1909.09223

Pombra, S. (2022). Analyzing and evaluating post hoc explanation methods for black box machine learning [Bachelor's thesis, Harvard College]. Harvard DASH. https://dash.harvard.edu/handle/1/37371734

Rudin, C. (2019). Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nature Machine Intelligence, 1, 206-215. https://doi.org/10.1038/s42256-019-0048-x

Rudin, C., Chen, C., Chen, Z., Huang, H., Semenova, L., & Zhong, C. (2022). Interpretable machine learning: Fundamental principles and 10 grand challenges. Statistics Surveys, 16, 1-85. https://doi.org/10.1214/21-SS133

World Bank. (2024). The use of alternative data in credit risk assessment: Opportunities, risks, and challenges. International Committee on Credit Reporting. https://documents1.worldbank.org/curated/en/099031325132018527/pdf/P179614-3e01b947-cbae-41e4-85dd-2905b6187932.pdf