Tribus attended the main RiskMinds conference last week in Amsterdam. The conference was a great opportunity to mingle with huge numbers of risk practitioners, consultants and academic heavyweights. One of the main topics of discussion during the conference – and especially during the third day – was the role of machine learning and artificial intelligence in model development and model risk management.
AI models for decision making – yes, but…
The consensus is that there is indeed a place for AI in model development to optimise processes and improve accuracy. Indeed, these models would enable modellers to tap into unstructured data to create more meaningful segmentations, identify new variables, identify previously unknown relationships and correlations, and ultimately create more powerful and accurate models.
However, there is also a consensus that this will expose firms to more and different risks. This view is primarily driven by regulatory requirements around the reproducibility and interpretability of models. Indeed, questions are being asked regarding the ability of these models to back-up a decision (especially if used in credit decisioning and the identification of suspicious money-laundering behaviour) and whether data lineage will remain robust.
In addition, the underlying biases and (sometimes lack of) quality of unstructured data, data privacy concerns (that unstructured data from the internet is as much your data as your bank statement), the self-learning potential of these algorithms and the difficulty in hiring skilled resources are some of the main challenges identified by presenters.
Regulators are starting to address these issues as they recognise their potential and attractiveness. In particular, some regulators have issued guidelines around the ethics, fairness and accountability of AI models.
MRM should “focus on what really matters”
A view emerged from the presentations that machine learning and artificial intelligence models can be used in model validation and more broadly in model risk management. Indeed, developing challenger models using AI, using powerful ML techniques to identify data quality issues and pre-populate validation reports are some of the areas of automation identified by presenters that would drive greater efficiency across MRM teams.
Having said this, the consensus was that MRM teams are facing a series of challenges that need addressing before ML and AI should be considered:
Ever increasing number of models in scope of MRM
A lack of a robust model inventory and cartography
Ever increasing regulatory and reporting requirements
Shortage of resources skilled in areas other than IRB models
Measurement of model risk, particularly for models other than IRB
Ever increasing cost-cutting across support functions
Nevertheless, exciting times ahead
The mood at RiskMinds was one of excitement at the opportunities that new ML and AI techniques can bring, but also one of caution as there is still a lot of work to do to ensure these techniques are safe to implement. After all, the viability of the financial system and, ultimately, the world economy depends on banks making their assessments as accurately and fairly as possible. There are many challenges ahead for model developers and model risk managers but also lots of opportunities. As risk practitioners, it will be our job to balance these out and move the industry forward.