Since the financial crisis of 2008, risk management has been a key area of focus for financial institutions. Faced with increased regulation and changing consumer expectations banks are constantly seeking ways better quantify and manage an increasing variety of risks. Banks need to keep capital productively employed, so there is a strong incentive to quantify risk accurately to minimize capital reserves required by regulators.
The financial crisis exposed inadequacies in financial risk management, particularly related to credit risk, where a borrower or counterparty is unable to meet obligations. Spurred on by new regulations, banks are now more attentive to a wider variety of risks including liquidity and insolvency risks, model risk, country, and sovereign risk, and operational risks brought about by failures of internal processes.
Augmenting financial simulation with machine learning
AI and machine learning are transforming the financial services industry. According to the McKinsey Global Institute, embracing AI and machine learning to develop new offerings, improve insights, and enhance risk management can generate more than $250 billion in the banking industry.(1) While banks are already embracing machine learning in lower-risk areas such as customer relationship management and digital marketing, they recognize that machine learning can improve the accuracy of risk models as well.
Machine learning models excel at identifying complex relationships in data, and banks are awash in data of all kinds – from customer profiles to transaction histories to market data to recorded conversations from call centers. At IBM, we see banks initially embracing machine learning to augment existing risk management techniques – essentially “checking the work” of operational risk systems and seeking to improve on them. Applications range from using Apache Spark and ML/DL to mine data to build better client behavioral profiles to using independently developed AI models to validate existing models and ensure regulatory compliance.
Machine learning carries its own risks
While machine learning holds promise, it brings new challenges as well. Machine learning models have shown that they can deliver better predictions in areas such as credit risk, but they tend to be much more complicated than their statistical counterparts. Because of this complexity, they are more susceptible to various forms of model risk. There are specific challenges related to interpretability and model bias.(2)
For example, a machine learning algorithm aimed at determining creditworthiness might recommend extending a loan to one client but refuse another. In this scenario, it may be difficult to explain clearly why a client was denied credit. Banks can use models to aid in decision making, but need to be careful about violating fair-lending laws.(3) If a model is used, it must be “interpretable” meaning it must provide clear reasons for refusals.
Machine learning models are also susceptible to various types of sample, data, and algorithmic bias. For example, anti-money laundering (AML) models using Random Forest algorithms(4) have proven to be prone to “overfitting” predictions to training data, placing disproportionate emphasis on model inputs with a large number of discrete values.
The risks in using machine learning and AI models
While the issues are complex, the good news for banks and regulators is that existing regulatory guidelines such as “Guidance on Model Risk Management” (SR11-7) issued by the US Federal Reserve(5) are considered broad enough to cover machine learning model risk. This means that machine learning models can be managed in the same fashion as established statistical and simulation-based models.
Risk Dynamics, McKinsey’s model validation arm(6) suggests six areas of model risk that banks need to consider as they embrace machine learning for financial risk applications.
- Interpretability – ensuring that AI-powered decisions can be explained
- Bias – avoiding sample, measurement and algorithm bias
- Feature engineering – ensuring the validity of model features
- Hyperparameter selection – having a sound basis for hyper-parameter section
- Production readiness – ensuring that models can be reliably deployed in production
- Dynamic model calibration – to continuously monitor and re-train models
A smarter infrastructure for financial risk management
While machine learning is on the rise, existing analytic methods aren’t going away. Clients will continue to use simulation-based approaches along with commercial and open-source analytic tools, including SAS®, MatLab®, R, and others. The challenge for these clients is how to extend existing environments to cost-efficiently support new ML-oriented frameworks such as TensorFlow, PyTorch, Caffe, and others.
Many banks already use IBM Spectrum Symphony software to accelerate and manage scalable risk applications. For these clients, IBM Watson Machine Learning Accelerator provides a logical place to start enabling an end-to-end workflow for machine learning. IBM Watson Machine Learning Accelerator provides complete lifecycle management for data ingest, data preparation, and building, optimizing, training, testing and deploying machine learning models. It also offers the advantage that clients can run new machine learning using the same on-premise or cloud-based grid computing infrastructure used for existing risk analytic applications, potentially avoiding substantial costs.
For SAS Grid Manager customers, the IBM Spectrum Computing Suite for High Performance Analytics enables them to run ML/AI applications on their current grid infrastructure. These IBM solutions can help pave the way to AI-powered financial risk models, providing essential capabilities needed to minimize model risk including feature engineering, hyperparameter selection, and ongoing model validation. Using this approach, banks can minimize infrastructure cost, manage model risk, and embrace machine-learning for financial risk applications at their own pace. And in doing so, they might help head off the next financial crisis.
- AI and machine learning generate value in the financial industry – https://www.mckinsey.com/business-functions/risk/our-insights/derisking-machine-learning-and-artificial-intelligence
- Machine Learning in the banking industry – a literature review – https://www.mdpi.com/2227-9091/7/1/29/pdf
- Fair lending in the age of machines – http://www.wolterskluwerfs.com/article/fair-lending-in-the-age-of-machines.aspx
- Random Forest and anti-money laundering – https://www.linkedin.com/pulse/random-forest-regression-tool-predict-alert-outcome-mayank-johri/
- Federal Reserve Guidance on Model Risk Management – https://www.federalreserve.gov/supervisionreg/srletters/sr1107.htm
- Risk Dynamics focuses on model risk management – https://www.riskdynamicsgroup.com/