10 Dec How Risk Managers Can Most Effectively Apply AI
Tremendous changes are taking place in the financial services industry. Data volumes are increasing, data availability is improving, information is being shared faster, and new methods of analysis continuously appear. Historical risk distributions have, consequently, become inadequate for assessing future risk. This is where artificial intelligence (AI) and machine learning (ML) come into play.
Though the terms are sometimes used interchangeably, ML is a core technique of AI that entails learning from available data and classifying and clustering it; the data analysis from ML can be used for robust forecasting – even for outliers.
AI, in addition to ML, includes automated decision making. It is a fashionable methodology that has many potentially helpful applications. However, each application must be matched with the appropriate AI algorithm, and each algorithm needs to be able to switch from historical data to more recent observations.
The most common uses of ML in financial applications involve big data and require instantaneous decisions. Marketing, high-frequency trading, retail credit risk assessment, cyber risk, know-your-customer and fraud detection are among the functions that can effectively leverage ML.
An example of a natural application of ML is when an angry customer calls, complaining either about an unauthorized credit card transaction or a transaction that was rejected by mistake. The AI algorithm can then incorporate this observation into its knowledge base, enabling better subsequent decisions to prevent fraud and approve or decline loans.
Indeed, any aspects of financial risk that deal with stable behavioral patterns are good candidates for AI. However, there is usually a trade-off between the speed of execution and predictive precision of the AI algorithm, on the one hand, and its transparency, on the other.
High-frequency trading or cyberattack prevention applications might use black box algorithms, as they must act instantaneously and with high precision. Many others, in contrast, require explainable AI (XAI). ML-driven models for credit scoring, loan approval and deposit pricing, for example, cannot function properly – and won’t be accepted – without understandable drivers and intuitive explanations.
There are ways, though, to find a trade-off between utilizing the best powers of ML and explaining it for verification and control. For example, regression-style ML can link various balance sheet and income statement segments – obtained by clustering deposits, loans or fees with similar characteristics – to macroeconomic or market variables.
When market trends change dramatically, regression-style ML algorithms – with regularization and cross validation – produce more robust results than regular regression. Such algorithms don’t require a priori hypothesis about which variables are suitable for the regression equation. Rather, they allow for a large number of potential explanatory variables, selecting the best ones automatically.
A common complaint about regression-style ML is that every explanatory variable must have intuitive justification, as data mining alone may pick up on random dependencies. On the other hand, a standard econometric hypothesis – which serves as the first step in regular regression – can miss the variables that are driving risks in a changing environment.
Regression-style ML algorithms help to expand a risk manager’s intuition about the kind of explanatory variables that are significant in unprecedented scenarios. Regularization and cross validation enable the algorithm to find an optimal balance between fitting the data (without overfitting) and having the best predictive power.
The important step here is the sensitivity analysis. Sometimes, excluding an unintuitive variable suggested by the ML algorithm delivers just a minor deterioration of the result, and is therefore acceptable. However, if the elimination of a variable creates a big dent in the outcome, it might be time to look further and find an out-of-the-box explanation. Human judgment should determine the final decision.
AI for Strategic Risk Analysis
Financial risk, especially during regime changes, can be compared to a specialized chess game, where, say, rules change all the time, the number of pieces is arbitrary, and the board size and shape are not fixed. So, it is critical to augment AI/ML algorithms with human oversight.
At the same time, the human mind cannot conceive of all possible combinations of events and their consequences. The job of AI is not to tell us what to do when it comes to strategic risks, but, rather, to extend our intuition with respect to understudied or unprecedented environments, so that we’re ready when they materialize.
For example, funded loan levels are usually positively correlated with the state of the economy. ML techniques capture this dependency, even though the correlation is not static.
Obligors plan higher growth, investment and spending when the economy is strong, and try to be more conservative when it slows down. But during really stressful periods, just before defaulting, obligors usually try to grab as much of unfunded commitments as they can, thus changing the correlation from positive to negative.
In many cases, financial institutions don’t have enough history of such dramatic regime changes to calibrate ML algorithms to account for this type of behavior. Experts, however, can make scenario-dependent adjustments to adequately capture this behavioral change, which drastically impacts exposure-at-default exactly when probabilities of default and loss-given default are much higher than usual.
There are several applications of ML-based classification and clustering algorithms that can predict loan delinquencies, but they are usually short-term (a month or a quarter forward) and focus on individual loans. Most significantly, these algorithms can (1) issue warnings to cut down unfunded loan commitments based on embedded covenants or other available considerations; (2) change approval for a loan extension; and/or (3) demand additional collateral.
Strategic risk management requires prediction of institution-wide loss when the correlation patterns between certain risk drivers change dramatically and there is a need to look beyond individual loans, modify risk limits (according to adjusted risk appetite) and possibly trigger contingency planning. Standard ML allows for such predictions, but sometimes with insufficient warning time.
The most effective way of dealing with such changing correlations is through generating the full distribution of macroeconomic and market scenarios that incorporate shocks, and then evaluating their effects on all variables. When balance sheet segments are linked to the scenarios using ML regression with regularization (as described above), the adverse scenarios can be automatically discovered.
Subsequently, algorithms like principal component analysis or spectral clustering can help identify the variables that have uniform behavior on the same scenarios. These variables often provide early warning indicators and contingency triggers. However, given the strategic nature of these steps, human controls, verification and sensitivity analysis remain an important part of the forecasting process.
The risk management community is still at the very beginning of exploring the applications of ML and AI. Every time there is an unprecedented market, we need forecasting tools that can look beyond historical data.
Right now, we’re facing some market-altering trends, including the transition away from Libor, the rise of negative interest rates and the necessity to measure the potential impact of climate change on financial institutions. This is where AI/ML algorithms, fed by growing data and combined with business experience and intuition, can provide insightful forecasts.