49 Finance Monthly. Bank i ng & F i nanc i a l Se r v i ce s conducted by the US Federal Reserve. CCAR requires a bank to submit its annual capital plans to the Fed to prove the bank’s ability to deal with severe economic shock. Given the sheer number of economic variables at hand, Citi’s increasingly manual modelling approach took too long to complete, leaving business leads little time to understand the logic behind the models. As a result, Citi couldn’t confidently defend its models in an annual submission to the Fed. To create an accurate revenue forecast model, Ayasdi started with enriching the macroeconomic variables stipulated by the Federal Reserve. Then, Ayasdi used its proprietary machinelearning software to reveal how exactly these variables impact the revenues of each business unit. This allowed the company to define which unit-specific variables have the most impact on revenues. The detected variables were then used to create a custom ML-powered model that can accurately predict business units’ revenues under stressful economic conditions. As a result, Citi has managed to make this compliance process three times faster and less resourceconsuming. Enhancing fraud detection Since the beginning of banking, fraud in its various shapes and forms has always been a persistent problem for many financial institutions. While many banks use sophisticated fraud detection systems, the rule-based nature of these solutions leads to a high probability of false positives. Importantly, fraudsters also keep innovating, which makes banks’ fraud detection systems grow obsolete. Machine learning models, on the other hand, can learn and evolve together with the fraudsters. In very simple terms, a machine learning model can detect unfamiliar deviations from the normal pattern, notify the human employee about it, and, based on human feedback, learn if this kind of deviation is the new acceptable pattern or a case of fraud. Just a few months ago, IBM launched a new generation mainframe that allows financial institutions to conduct fraud analysis of 100% of their transactions in real-time. To compare, around 10% of transactions could go through an AI-based fraud detection engine. In a nutshell, this means that banks that use the z16 mainframe can significantly reduce the number of false positives and increase customer satisfaction. “While many banks use sophisticated fraud detection systems, the rule-based nature of these solutions leads to a high probability of false positives.” Conclusion Risk management is a natural playground for machine learning. With the amount of data that banks have accumulated in the past decades, it’s only right to use the technology for its analysis. Machine learning models can drastically increase the accuracy of forecasts, decrease operational costs, and make risk management more effective overall.