In this lecture, we'll talk about additional applications of machine learning in finance beyond just fraud detection. In the conversation with Apoorv, we talked about some of these examples, but in this lecture we're going to go into the details of how we can apply machine learning for a variety of different needs within financial services. To begin with, let's talk about identity verification and authentication. This is extremely important in financial services because if you're allowing consumers to transact or to move money around, we want to make sure that the person who's actually initiating the transaction is who he or she claims to be. We need to do identity verification. The traditional way of doing verification of identity is by using passwords are using PIN numbers. But the application of machine learning would be in terms of biometric identification, for example, to do face recognition or to do fingerprint recognition, or to do voice recognition in order to identify who's actually interacting with the company on the website or on a mobile app. Now, in terms of these applications, you can imagine a situation where a customer is trying to log into their bank account on a website or on the mobile app, and we might ask the person to take a photo of themselves as part of the login process or speak into the device, or perhaps use their fingerprint to try and log in, so these are some examples. In fact, in banks in China, it is not uncommon to have ATMs use face recognition. When a customer walks in front of the ATM, a camera that is on the ATM might actually recognize the customer, then authenticate them and allow them to use the ATM. This is in fact being used by multiple banks in China today. Bio-metric identification needn't just focus on things like a person's fingerprint or their voice, over time companies and researchers are also testing other ways to do customer identification, such as look at the unique ways in which they hold the phone, or the unique ways in which people tap the phone. This technology is not mature enough. It's not clear whether it will actually be used in practice. But today there are multiple uses of biometrics for identification, and all of them tend to rely on machine learning to do the identification. The advantage of these kinds of techniques is, of course, the added security that biometrics provides relative to things like passwords. Of course, a limitation is that they're not foolproof, they do not guarantee that a fraud is not occurring, but certainly they can help minimize it quite significantly. Another application of machine learning is in loan and insurance underwriting. If you look at loan approval decisions, it's made today by loan officers based on some information about customers. But increasingly there is interest in applying machine learning algorithms to either guide or make the loan decisions themselves. These algorithms tend to typically be supervised machine learning algorithms, so they get a training data set. The training data set would include a number of features or covariates or x-variables, which is customer data, such as their age, their income, their employment, their past history, their credit rating, and so on. It will also include a cleanly labeled outcome variables such as what is their credit score in the past or what is their history in terms of repaying loans on time or whether they've defaulted or not. Supervised machine learning algorithms can then analyze the data and figure out what factors are predictive of whether a person will pay back on time and then make the loan approval decisions. There are many advantages of these approvals. For one, they can help speed the loan processing time and increase the loan volume that is handled by banks. They can also use more diverse data sources and help make loan approval decisions for groups of applicants who might not have rich loan history. Young people as well who don't have a credit history, and so that can be applied in those settings as well. Of course, a limitation here is the algorithms can themselves end up being biased because they are based on past data, so if there are biases in historical data, for example, minority groups might not have been given loans in the past, or women might have been discriminated against, some of those biases can enter the algorithms. One has to be careful about that. Now having said that there can be biases in these algorithms, I want to note that it is important to recognize that these biases come because there have been biases in the past by humans. When we become nervous about AI biases, we have to recognize that the alternative of AI is a human being and the human might themselves be biased. Also furthermore, it is possible to run statistical tests and detect biases in these algorithms and hopefully correct them at scale. The point here really is that there can be biases. We have to be careful about them, but they can also be detected and improved, and so companies should not be hesitant to use AI just because of the risk of bias. They, in fact, should be more proactive about detecting and testing for biases and correcting them. Now, another application of machine learning within financial services could be in detecting likely churn customers. That is detecting which customer is likely to leave you based on the patterns of usage, based on things such as the number of times they log in to the mobile app, the number of times they might visit the store or the kinds of transactions they engage in. This might help a company, and it could be a bank, but it could be any company, even outside of financial services. They can detect who's likely to churn and then take preventive action, such as engaging the customers through special offers and discount. Now the benefits of this are very obvious. If a company can detect churn and prevent churn, then the lifetime value of the customer increases and the enterprise value increases quite dramatically. Of course, a limitation is that machine learning can help detect churn, can help identify the factors that are driving that prediction, but that doesn't mean machine learning can tell us what is the exact action for us to take in order to retain this customer. That's where managers like human comes in, and so we should think about the machine learning as providing very valuable inputs for human decision-makers to then intervene and take important actions here. Lastly, I want to mention that these are just a few examples of using AI within financial services. There are in fact many others Apoorv Saxena, now the Global Head of AI at JPMorgan Chase talked about how conversational AI platforms can be used to interact with customers via chat or over the phone to improve responsiveness and reduce costs, especially in conversations that are high frequency and low touch. He also mentioned personal finance, for example, creating personalized portfolios for individuals. That's another interesting area. Financial forecasting, the ability to predict, what kinds of financials a company might have in the future or budgeting needs. This is another area. In short, there are many different areas. There is certainly no scarcity of options or applications of AI in finance. The important issue becomes, how do we prioritize them? How do we actually determine which ones to focus on first and which ones to focus on later, what kinds of resources to put in place, and so on.
