AI in Banking
Customer Churn Prediction
AI is a great solution for customer churn prediction as the problem involves complex data over time and interactions between different customer behaviors that can be difficult for people to identify. AI can look at a variety of data, including new data sources, and at relatively complex interactions between behaviors and compared to individual history to determine risk. AI can also be used to recommend the best offer that will most likely retain a valuable customer. In addition, AI can identify the reasons why a customer is at risk and allow financial institution to act against those areas for the individual customer and more globally.
Banks and credit card companies use credit scores to evaluate potential risk when lending money or providing credit. Traditional credit scoring uses a scorecard method which weights various factors including payment history, dept burden, length of credit history, types of credit used, and recent credit inquiries. This traditional method is based on broad segments and will deny credit to consumers without considering their current situation or other extenuating factors. Traditional methods may also give credit to consumers, called churners, who are “gaming the system” and taking out a large number of reward credit cards but are not profitable for the issuers. For credit decisions there is also the additional regulatory burden that banks and credit card companies must explain to the consumer why they have been denied credit.
AI, especially time series modeling, is particularly good at looking at series of complex transactions and finding anomalies. Anti-money laundering using machine learning techniques can find suspicious transactions and networks of transactions. These transactions are flagged for investigation and can be scored as high, medium or low priority so that the investigator can prioritize their efforts. The AI can also provide reason codes for the decision to flag the transaction. These reason code tell the investigator where they might look to uncover the issues and help to streamline the investigative process. AI can also learn from the investigators as they review and clear suspicious transactions and automatically reinforce the AI model’s understanding to avoid patterns that don’t lead to laundered money.
Know Your Customer
Know Your Customer is a key part of money laundering and anti-terrorism legislation. The Customer Due Diligence (CDD) process requires banks to file reports of suspicious activity. Almost two million such reports were filed in the United States alone in 2017 according to a study by the Royal United Services Institute for Defence and Security Studies — a U.K. think tank. Failure to identify and file reports on suspicious transactions results in billions of dollars in fines for banks. Investigators looking into suspicious activity use a variety of tools including rules that flag frequent or international transactions or interactions with offshore financial centers. Unfortunately, with the volume and variety of transactions, rules-based approaches are not flexible enough to capture new patterns and produce large number of false positives that need to be reviewed.
AI can be used to analyze large volumes of transactions to find fraud patterns and then use those patterns to identify fraud as it happens in real-time. When fraud is suspected, AI models can be used to reject transactions outright or flag transactions for investigation and can even score the likelihood of fraud, so investigators can prioritize their work on the most promising cases. The AI model can also provide reason codes for the decision to flag the transaction. These reason codes tell the investigator where they might look to uncover the issues and help to streamline the investigative process. AI can also learn from the investigators as they review and clear suspicious transactions and automatically reinforce the AI model’s understanding to avoid patterns that don’t lead to fraudulent activities.
AI systems have been proven successful at detecting anomalies in transaction volume data. This time series process looks at expected data volumes based on historical patterns. Upper and lower boundaries are also predicted based on volume variation. This system is then used to compare real-time transaction value to expected volume. This real-time system allows network administrators to be notified when transactions start to spike above or fall below these boundaries so they can take action before an outage in service.