Machine learning in credit risk was moving from experimentation toward broader application across financial institutions, according to a 2019 survey of 60 firms. The findings offer a useful view of how adoption, use cases, and implementation barriers were evolving at that time. For technical and operational leaders, the central lesson is that successful deployment depends on more than model performance: data, infrastructure, governance, and clear communication all matter.
Machine learning adoption moves beyond early pilots
The survey found that 42% of participating institutions had machine learning models in production, while a further 45% had active pilot projects. Another 10% planned to begin using the techniques within six to 12 months; 3% reported no foreseeable plans to adopt them. Compared with the previous survey, pilot activity had increased sharply, while production adoption had risen more modestly. Among institutions already using machine learning, the sophistication of models had advanced.
Adoption was not confined to large firms or developed economies. The report instead associated maturity with each institution’s business strategy and innovation agenda. It also found increasing activity across regions, including a notable rise in Japanese pilot projects compared with 2018.
Where machine learning is used in credit risk
Credit scoring and decisioning remained common applications. However, institutions were increasingly applying machine learning to other parts of the credit risk process, including credit monitoring, early-warning systems, collections, restructuring, and recovery. The survey described a shift away from regulatory uses such as capital, stress testing, and provisioning, where requirements for simpler, interpretable models could make some machine learning approaches harder to apply directly.
Institutions varied in both the scope and complexity of their work. Some applied machine learning across multiple products and customer segments, and across activities such as data preparation, segmentation, model development, and validation. Others focused on a single segment or used machine learning for specific tasks, such as variable selection. Techniques also ranged from using machine learning within part of a model-development process to building full machine learning models, sometimes incorporating unstructured data.
Growing interest in SME portfolios and early warnings
Retail portfolios remained an important area of use, in part because firms often had larger volumes of standardized data for these customers. At the same time, production use in small and medium-sized enterprise (SME) portfolios increased substantially between the survey editions. This expansion suggests that institutions were extending applications beyond established retail use cases.
Some firms also used natural language processing to analyze material such as news, annual reports, supply-chain information, and social media. In the surveyed applications, these sources could support early-warning models intended to identify signs of deteriorating credit. The report also noted use of machine learning for model validation, including the development of benchmark or “challenger” models to compare with existing approaches.
Benefits: prediction, data analysis, and risk segmentation
Respondents identified improved model accuracy as a key benefit. They also cited the ability to work with data that is incomplete or inconsistent, uncover previously unnoticed risk segments or patterns, and analyze more variables. By considering information across a broader set of inputs, machine learning can help institutions explore relationships that may be difficult to identify with more conventional approaches.
The report described these capabilities as potentially useful for developing a more granular understanding of different population segments. That analysis may help firms evaluate customer groups that were not well served by earlier models. However, the reported benefits do not remove the need to assess model suitability, data limitations, and outcomes in the context of each application.
Key challenges: explainability, data, and infrastructure
As institutions gained experience, they reported a wider range of challenges. Supervisory understanding or consent for new processes, difficulty explaining model behavior, IT infrastructure limitations, and access to appropriately skilled staff were among the prominent concerns.
Explainability was important to colleagues, auditors, and supervisors who need to understand how a model produces its results. The report noted concerns about transparency, auditability, and interpretability, particularly for models perceived as “black boxes.” It also described ways institutions were working through these concerns, including expert judgment, validation tools such as partial dependence plots, and closer examination of input data. The survey did not suggest that one method resolves every explainability question; expectations and approaches can depend on the use case and supervisory context.
Data quality presented a related but broader obstacle. Multiple sources and formats, as well as missing or inaccurate information, can complicate modeling. Respondents noted that these issues affect traditional models too, rather than being unique to machine learning. Legacy IT environments can add further friction: systems developed incrementally may be siloed, redundant, or difficult to use for data aggregation and analysis. Some older systems may also lack support for modern coding languages, requiring translation work.
Finally, institutions reported difficulty finding people who combine machine learning, finance, and coding expertise. The report described internal upskilling as one response to competition for external talent. Skilled users remain essential to selecting, developing, and evaluating models responsibly.
Practical implications for model deployment
Regulatory engagement increased among institutions with production models. Interactions were often described as constructive, although some firms wanted clearer guidance, including on the level of explainability expected before deployment. The report emphasized the value of collaboration between financial institutions and supervisors, so that innovation can be assessed alongside customer protection and risk considerations.
For teams evaluating machine learning in credit risk, the findings point to several practical priorities:
- Define the intended credit risk use case and determine whether the model’s complexity fits its purpose.
- Assess data completeness, consistency, sources, and the capabilities of the supporting IT environment.
- Plan how results will be validated, explained, audited, and communicated to relevant stakeholders.
- Engage supervisors early where new processes or production models raise questions.
- Build the skills needed to develop and oversee models, including through internal training where appropriate.
The 2019 survey captures a period of rising adoption, not a guarantee that every application will deliver the same results. Its practical message remains clear: machine learning can broaden analytical options in credit risk, but production use requires coordinated attention to model performance, data quality, explainability, infrastructure, governance, and people. These foundations help institutions evaluate innovation without treating the model itself as the whole solution.
Access Official Technical PDF