EP 133: How AI Will Change Financial Risk Management

This is the Everyday AI show, the everyday podcast where we simplify AI and bring its power to your fingertips. Listen daily for practical advice to boost your career, business and everyday life.

How will AI change financial risk management? You know, and how, how are financial institutions going to be impacted with the rise of AI and what does that ultimately mean for us, the consumer? All right, that's what we're going to be going into today and more on Everyday AI. Welcome. My name's Jordan and this is your daily podcast, free newsletter, live stream, helping everyday people like you and me make sense of what's going on in the world of AI and how we can actually use it to grow our companies and grow our careers. So we're going to be diving into what's happening in the world with financial risk management and how AI is impacting all of this. It's super interesting because, you know, AI has been used very widely for many decades in the financial institution, but recent kind of updates and features, I guess, in generative AI are going to be impacting financial institutions as well. So very excited to get into that. But before we do, let's go over the AI news. And as a reminder, if you're joining us live like we have here, Dr. Harvey Castro joining us live, Christy Slack joining us live. Thank you. Get your questions in. What do you want to know about how AI will change the financial risk management sector? All right, let's get going. Go over AI news. We're actually just going over two pieces today because they're bigger pieces, but as always, there's more news. Go to your everyday AI.com Sign up for the free daily newsletter. All right, so the White House has released the US Government's first ever executive order on AI. So President Biden unveiled a new executive order on artificial intelligence that aims to address safety concerns, protection of civil rights, and support for workers in the industry. So this new executive order involves creating new standards, protecting consumer privacy, promoting innovation and competition, and also collaborating with international partners. This executive order is the first binding action ever taken by the US Government on artificial intelligence and includes regulations for large companies to share safety tests with the government before release. That's an important part. We'll see if that happens. Also, this prioritizes the development of AI standards for testing and watermarking, as well as guidelines for agencies using commercially available data. So pretty big announcement again. We covered this last week when the news came out, but this executive order was just released and we're going to have a lot more on this one in the newsletter today.

All right.

Our second piece of AI news and last one for today, some new chat GPT updates. All right, so nothing official yet from parent company OpenAI but if you've paid attention at all over the last two or three days over the weekend to the Internet, to social media, you'll see that ChatGPT is already starting to to unveil some pretty big updates. So many users are now sharing these new updates and probably two or three of the bigger ones are a reported updated knowledge cutoff date of September 2023. So this has already changed once. It went from September 2021 till January 2022 for GPT4, the paid model. But apparently now it's be gonna being rolled out. The knowledge cutoff is all the way up to 2023 as well as kind of the other big feature or the big announcement here is the all tools mode, which is essentially being able to upload photos, PDFs, receive dolly images all inside one mode. So without having to flip back and forth between multiple modes. So really getting the first taste of multimodal in one single chat versus having to go into multiple different modes. All right, a lot more on that later in the week. We're actually going to have a dedicated show on new chat GPT updates and the future of chat GPT with the developer conference coming up this week. All right, but you didn't hear, you didn't tune in to hear about ChatGPT and you can again always go to your everyday AI.com for more on that. You are here to learn about how AI will change financial risk management. So I'm very excited to have today on our show and please, please help me welcome. We have. There we go. We got him, got him on the screen now. Sandeep Myra, the founder and C. CTO of Raven Risk Intelligence. Sandeep, thank you for joining us.

Thank you so much for having me. Really, really honored to be on the show.

Yeah, absolutely. Let's, let's start high level real quick. Just tell everyone a little about yourself and about what you're doing at Raven Risk and Intelligence.

Yeah, so I'll keep the part about myself pretty brief. You know, my background's in computer science. I actually took an AI class at Cornell 30 years ago. It's kind of remarkable when you think about the. What's happened, you know, recently versus the last 30 years and one. So I side anecdote was in that class. I took out the textbook from the class, I still have it a few days ago and it said there's a small section neural networks which is what LLMs are based on and ChatGPT is based on. And it said it's not showing much promise yet, but the people, the neural network researchers think that with enough computational capacity and data, it's going to emulate human decisioning. And it said only time will tell actually said in the textbook only time will tell what actually happens. So anyway, so basically, I think there was a long AI winter. I mean, in the meantime, while I had an interest in AI more broadly, had an interest in analytics, I worked with a lot of financial firms and applying essentially, I would say algorithmic techniques for financial risk management in particular and some trading, you know, systems. So I've worked at J.P. morgan, Citigroup, BNY Mellon, and more recently founded earlier this year an AI venture called Raven Risk Intelligence. And, you know, I'm happy to talk a little bit about the objectives of the, you know, of the venture. Jordan, if you want to meet, Want me to go there next?

Yeah, yeah, let's, let's just go high level. Let's just talk about a little bit about what it is so everyone can understand. So, yeah, just tell us a little bit about that.

Okay, so essentially, and I think very broadly, the three thesis is twofold. So one of them is the ventures actually in commercial and corporate credit lending, not the consumer credit space and the commercial credit space, there's a lot of manual effort done by large teams of credit analysts to try to gather information about the borrower and also the economy. So it includes things like finding out about the business strategy of the company, the strength of the management, you know, any competitive threats to the company. And then from a more macroeconomic standpoint, it would be, you know, things that might be happening in the economy and so forth. Now, today, basically, most of the information, automated information that they get tends to be pretty static. You know, they have to actually kind of scour essentially unstructured data sources like the ones that I'm just mentioning to come up with a view on whether this is actually a good credit or not. So our first goal is to help automate that, which will lead to actually increased ability to process more loans and frankly be able to give more loans to more companies by taking into account a broader set of inputs rather than just is the company profitable today or not? And that actually, frankly, I think will enable a lot of smaller borrowers to get loans more easily. And the second part of it is what we're calling predictive risk analyses, broadly speaking, which is, you know, how are things going to perform over, you know, like a wider period of time, but using fairly advanced analytics and machine learning analytics to draw correlations between, you know, different things that are happening in the, in the industry and in the economy.

I love it. And Cindy, maybe help us also, you know, for, for those of us that don't follow the financial sector very closely. You know, let's just talk, you know, briefly about you know, financial risk and, and risk management and, and kind of historically, you know, where, where it's been recently and how you see it changing now with advancements in generative AI, you know, and like, like you talked about, you know, I love that you mentioned the, the textbook from, from 30 years ago and you know, kind of the AI winter. But now we're getting to the point where, you know, AI is really helping in that decision making process. So like, what does that mean broadly for the financial risk management industry?

Yes, I think, you know, like very, I think broadly speaking, you know, some of the things that I think you. So let me just back up a little bit the way that, you know, risk management works today. And even actually the more advanced risk management techniques, you know, tends to be taking into account what we call structured data. And that data is, you know, things that are tabular and format like you know, rows and columns. So things, you know, know, simple stuff frankly, like, you know, what is the revenue of the company, you know, what essentially they do take into account some macro economic indicators like, you know, what is the GDP growth in the economy, etc, you know, and I think, and that's the, that has been the cutting edge actually unbelievably like that's the limit essentially of what their, their automated sort of tools and risk management can do today. And then they take those, what I call structured and the industrial structured risk factor, structured input, sorry, and then, you know, put them into models that attempt to make decisions about or outlooks of risk for a given, let's say company or sector. You know. Now the problem with that, which is kind of alluding to earlier, is that it's actually relatively narrow. It doesn't take into account things that, you know, let's call it unstructured, which are information pieces that might, you know, happen that are not, you know, tabular in format, frankly. So an example could be that, you know, there's a war that breaks out somewhere, you know, like maybe in Taiwan or wherever. And you know, these models actually can't really take that into account at all. You know, it's human judgment that then tries to figure out, oh well, what exposure does this company have to Taiwan and you know, and that can be a very manual and B, not very comprehensive, leading to frankly, inaccuracies. So I think the objective is that, you know, with the event of machine learning, you could take these events that happen and basically in real time, which is pretty amazing, and then, you know, correlate that with impacts to, you know, to the economy and to even individual companies.

Yeah. And just real quick, and maybe, Sandy, if you can even help, help me better understand this because I'm always trying to learn as well. So with, with unstructured or, sorry, with structured data that's been used in the, you know, financial industry and for risk management for decades. Right. So that's where, you know, machine learning and AI gets all of these data points and they can categorize them. Right. And they can say, yes, this pattern of data over the course of, you know, hundreds of thousands or millions of data points, we can make decisions based on this structured data. Whereas unstructured data, it's a little harder for AI models to be able to understand that and to be able to translate that to risk because it could be things that require more interpretation or more interconnectivity that may be hard to, for traditional AI models to perform those tasks. But that's maybe where now with large language models where you can start to make use of some of this unstructured data and tie it to risk management or to assess risk. Is that kind of a good overview? And then if so, how do you think large language models might be able to help pull this all together?

Yeah, well, firstly, I think that's an incredibly perceptive observation. Frankly, I was worried.

I was worried.

I wouldn't say that you don't know a lot about how to use AI and risky program probably actually know quite a lot because I think, you know, you've connected a lot of the dots actually, which is what these models, you know, obviously are doing in terms of trying to figure out what, you know, the impact is to, to risk. So yes, I mean, I think, you know, the, you're correct, firstly that structured data and having models, even some machine learning models actually draw correlations on structured data has been around for, you know, some, some years and you know, they've done a pretty good job actually. So for example, in fraud detection, you know, whether it's credit card fraud or, you know, even a trading fraud, you know, these models have been around for few years where they look at different patterns of behavior of, let's say, consumer borrowing. So let's say that, you know, you go basically abroad somewhere that you haven't been before. You know, you'll notice quite often that the credit card company will, you know, call you up or even block your card from usage because they're noticing, you know, essentially an anomaly in your, you know, in your, in your credit behavior. So that's been around for a few years. But I think what is new though is to, you know, use essentially this for other use cases and do it at a much larger scale. And so an example is, you know, Silicon Valley bank actually might be a good example. So in Silicon Valley bank, you know, what happened was that the, you know, the Fed raised interest rates very rapidly. The bank essentially had what's called a liquidity. So that's called market risk. You know, rates are considered to be like market events that led to what is called liquidity risk issues, which is that the bank didn't have enough money, cash on hand to actually satisfy all its depositors. Because banks take depositors money and loan them out. They're not just sitting in the bank because they have to earn interest for the bank so that they can pass that interest onto the depositors. So they had what's called a liquidity risk issue. And because of that they had what is called a credit event, which is the bank essentially, for practical purposes, defaulted, which is essentially meaning that they could not satisfy their creditors, who actually their depositors are the creditors in this case. I think that interconnectivity would have been much more easily apparent with the use of proper training of AI models and how different risks are interconnected to each other. And I think what happened at Silicon Valley bank would have been almost completely predictable with the better use of this interconnected AI models that I think you're talking about. And large language models in particular, you know, just to double down on that part a little bit, are actually really good at that. So they basically, you know, I know they call large language models, but underneath the covers what they're doing is looking at, you know, connectivity and correlations between different things and then figuring out essentially, you know, what to so called generate. And that's why it's called generative AI. But you can use that not only just for pure language, but you can actually use it for drawing correlations and patterns essentially between all kinds of data sets that was not achievable before. So I think those large language models can be very, and those techniques, I guess the modeling techniques that are used in LLMs can be very useful for risk analytics as well.

Yeah. And hey, as a reminder, if you're just joining US live. Midway through, we have Sandeep Myra, the founder and CTO of Raven Risk Intelligence. And if you have questions, please get them in now so we can give Sandeep a question, a chance to answer your questions. And Sandeep, I'm so glad you brought up, you know, this Silicon Valley bank kind of collapse because I think that's maybe one of the most relatable for many people in terms of financial risk because we saw unfortunately things go down in unfortunate path for, for many involved. And I think some of the initial response to that is people said, hey, you know, with, with all of this data, with all of this, you know, artificial intelligence and machine learning, how did this happen? And you, you kind of started to, you know, help us solve that. It's, you know, kind of different, I guess, models or different sets of data that maybe weren't talking to each other. So, you know, with, with even generative AI, I guess that could potentially help solve this in the future. What are still those obstacles to overcome until we can have, you know, generative AI, you know, help kind of, you know, quote unquote, connect all these different, you know, pieces of data or these different models together? What, what do we still have to do? And then maybe even what are the risks of doing that?

Yeah, no, that's a great question. So, you know, so firstly, I think, you know, these models can, are only as good as the data and how the data is essentially presented to them. And so, you know, they're not magic. I mean, they basically might seem like magic, but the reality is that all even ChatGPT is doing is it's taking all the data on the Internet and, you know, trying to do its best essentially to come up with what makes sense from an output perspective. But not everything is on the Internet. So particularly I think in some of the, you know, business domains, like in commercial credit lending, you know, a lot of the, you know, a lot of essentially the inputs actually come from human inputs that are not codified on the Internet. You know, so for example, like, you might have something that is somewhat subjective about, you know, essentially let's say that, you know, there's this going to be a change in the business strategy of the company, as an example. And then quite often there's a subjective decision made by the bank about, you know, does that business strategy lead to potential risk the company or not? And, you know, and what, and how big is that risk? You know, like, and I think so the, these large language models are not yet at the stage where they can codify Things that are and even come up with correlations for things that are not, you know, readily present the data. And so human, what's called human reinforcement learning, there's actually a couple of terms for it. One is called long, long winded term. I mean, these guys come up always in the space with very long acronyms and you know, unobscure terms. But it's called reinforcement learning with human feedback RLHF. And that actually is actually a pretty hot area of even research, ironically, is to actually get humans to at least partially train the more obscure and more critical parts essentially of these models. Because the impact of these models and business decisioning can be pretty severe. So somebody could be denied a loan, for example, and you know, could actually mess up their business if the models present data that is, you know, outputs that are not, not completely accurate. So that's one thing. So that's, I think one thing that I think is, you know, like a challenge. But I think there's some ways, like I said, it's an active space is to not to take these Automated models and LLMs and chat GPT like models, but then in, you know, apply some human oversight and inputs onto that, onto the modeling process. The second one is, you know, that in finance, and particularly in regulated finance like banks, you know, they tend to be very highly regulated. And so the regulators are very nervous about using machine learning, frankly, in general for decisioning purposes. They started to get somewhat more comfortable about using machine learning for using structured data, particularly for things even like consumer credit. But they're not yet there in terms of using unstructured data and machine learning to come up with decisioning. I think we all know about, you know, many of us know about hallucinations. So these models are not completely accurate. You know, quite often, frankly, if you're, they're asked very specific questions, you know, they can. Some of the data is inaccurate, which is frankly not acceptable in the finance space. I think the regulators are very nervous about that and probably rightly so. So I think one of the things that, you know, I think is going to be a challenge is actually a getting the models to be more accurate than they are today. So moving them from a consumer space to an enterprise decisioning space and then getting, you know, once that happens and getting the regulators comfortable, which is not always easy, with hopefully improvements and you know, in the, in the decision recommendations from these models. And then a related point to that actually is something called explainability is that the regulators and even the firms themselves, you know, don't like black boxes so they want to know essentially some idea of how the models came up with these outputs and recommendations based upon the inputs. You know, so they want some traceability between the inputs and the outputs. And you know, unfortunately today deep neural networks like ChatGPT are, are not able to do that. I mean they're, the models are so large that it's not easy to actually trace how chat GPD came up with the outputs based upon, you know, billions of points of input from the Internet. So that's going to be another frankly challenging area as well.

Are you still running in circles trying to figure out how to actually grow.

Your business with AI?

Maybe your company has been tinkering with large language models for a year or more, but can't really get traction to find ROI on gen AI. Hey, this is Jordan Wilson, host of this very podcast. Companies like Adobe, Microsoft and Nvidia have partnered with us because they trust our expertise in educating the masses around generative.

AI to get ahead.

And some of the most innovative companies in the country hire us to help with their AI strategy and to train hundreds of their employees on how to use Gen AI. So whether you're looking for ChatGPT training for thousands or just need help building your front end AI strategy, you can partner with us too. Just like some of the biggest companies.

In the world do.

Go to your everyday AI.com partner to get in contact with our team or you can just click on the partner section of our website. We'll help you stop running in those AI circles and help get your team ahead and build a straight path to ROI on gen.

You know, so as AI can help in all of these areas and I'm sure that's where the, you know, the financial risk management, you know, experts are starting to spend their time in. Where do humans kind of fall into this, this future equation, right? Like is their job going to, to change, are their responsibilities going to change? And, and like I said, is this maybe a good thing or a bad thing? And, and, and you know, what are even the risks of, of that? As, you know, leaders working in this space are maybe using and leveraging more, more AI. What do, what do we have to keep an eye on to make sure that this is successful in terms of, you know, risk management and handing these things off? Because it sounds like AI can really help in some of these areas and to, you know, help connect some of these disjointed, you know, verticals where we have all this different data that exists. But then, you know, how does that change? Then, you know, what ultimate responsibilities, you know, lie on us, us humans.

Yeah. So basically, you know, firstly I think in terms of the, the, you know, things related, I think people are nervous, you know, certainly. And I think frankly the more people use chat GBT in some cases, the more nervous they get because they see, you know, how powerful it is. Right. So people, you know, for, for good reason, you know, well, we're worried about their jobs. I even get questions from people about, you know, what fields their kids should study in, you know, that will be less adversely impacted by AI, you know, so there's a lot of, I would say valid actually questions and concerns about, you know, the impact to, you know, to social impact, but also, you know, impact to the workforce. You know, my view is somewhat more, somewhat fairly positive, at least in the long run. So you know, I think in the long run things that are more tedious will be taken away. And you know, essentially the machine learning and AI can automate those tasks. But there are things that I think are harder actually for machines to be responsible for that humans actually could play a bigger role. So as an example is even in this risk space and the productivity that I was talking about, I was talking to somebody very senior at a big bank and risk and he was saying that he's found that their credit risk analysts actually find this gathering of information and then trying to summarize it to come up with some outlooks is very tedious. And actually the turnover he's found in that part of his team is actually pretty high. So I think essentially it will remove some of the more tedious tasks and enable humans to focus things frankly that are more interesting and more value add. So I think there are things that we can do in the future that we don't even know yet. I mean an example could be in the media space that people are very worried about AI basically generating movies automatically and taking away actors jobs. Now in the short to medium term that's a valid concern. But in the longer term, if you think about it, somebody who's very creative could essentially, as one person potentially, let's say eventually in the future, create a full length movie on their own, which today is very difficult for creative people frankly to break into getting large audiences. It's not an easy task. I do think that there's opportunities for leverage here. If you think about it. Just one last point on that. It's not completely dissimilar to the Internet, you know, where essentially people were very worried that the Internet would take away lots of jobs, particularly in some sectors like Retail when Amazon, with the advent of Amazon and so forth. But I think if you look at another way, there were many jobs created related to the Internet that, you know, in many ways offset the, more than offset the job losses in other sectors.

You know, one thing, one thing I, as a super aside, one thing I can't help is me charging my, my mouse battery. So apologies. I do see, I do see some great comments coming in, but my mouse actually died so I can't bring them up. I'm sorry. But maybe Sandeep, you know, as, as we look, as we look forward to the future of risk management. So one thing I, I maybe want to get, get your thoughts on is how this ultimately impacts consumers. Right? Because I think, you know, if, if we're su. Like if, if, if we're looking like what's actually very tangible to consumers, you know, one thing that we probably worry about is, you know, risk and fraud. How might we, the average, the average, you know, bank consumer, you know, we have our Savings accounts, our 401ks, our IRAs, credit cards, all those things. How might we be impacted by all of these changes that we're kind of talking about? And even as it comes to risk, you know, are consumers ultimately more at risk in the long run? Or maybe are we less?

No, I think actually in risk it's a net, and frankly a net positive. Because, you know, like, I think one of the issues in the consumer space today is that people who don't have what's called a traditional credit history find it hard to get, you know, credit, including credit cards, loans. And you know, I think by taking essentially what we're calling unstructured data sets, like for example, let's say that somebody doesn't have a long credit history, but they've got a, you know, a good history of paying their regular bills on time, like the utility bills, etc, Right. And they, they've been, you know. So I think those kinds of data sets that haven't been used today could provide essentially better, you know, like I would call outlooks in terms of what the consumer's ability or ability to pay back essentially the credit looks like. So I think it actually will expand, in fact, access to credit for consumers who have had a harder time getting credit today. And frankly that includes minority populations or people who through no fault of their own, have had, let's say, a rough time. But inherently they probably can be a good credit going forward. So I actually think it's a net positive. On the commercial space, the impact is probably less directly visible. But Another way to look at it is that if companies get easier access to credit to that essentially helps these business owners, some of them are small business owners as well, not just large corporations. And then that ultimately helps the economy by ensuring that the economy is more productive and reduces prices for consumers. So you don't want an economy where the access to credit by both large and small business businesses is appropriately allocated because that ultimately drives, you know, what benefit and prices that consumers pay, you know, on the street.

What's. So Cindy, we covered, we covered an awful lot here. You know, we talked about, you know, historical use cases for AI and machine learning over many decades. Right. Like even going back to, you know, a course you took some 30 years ago and then of got caught up to, to current day and you know, some of the challenges and also some of the opportunities that, that are associated with, you know, financial risk management in the new age of, of generative AI. But maybe, maybe what's that one, that one point that you would really want to stick with people, right? So whether they're in the financial industry or if it's just everyday person, what's kind of that one big takeaway that you would want us all to, to hopefully understand so that we can better understand kind of where financial risk management is going now, now that we have access to better, more powerful and more, more connected AI systems.

Yeah, I think the general point that I have, and maybe it's not just specific to financial risk management and it's maybe somewhat obvious perhaps, but is that, you know, I think that everybody should stay current with, you know, the tools that are out there, like chat GPT. There's another way to think about it is that if you're not using those tools, then, you know, your counterpart may be using their tools and improving their productivity. Right. So like, you know, we, we want to, you want to be a little bit careful from your own, I would say, career perspective as an example that you're staying current with what the openly available capabilities are for like, you know, large language models as an example because you don't want to be put at a disadvantage, right? You want to, you know, my strong recommendation is to stay current with what's happening with at least the widely available tools so that, you know, you can use them. Obviously we're allowed, you know, to improve essentially both your personal life as well as maybe your productivity at work. So that's that. I think what I would say to, you know, most people.

Sound advice. Sandeep, thank you so much for, for joining the Everyday AI show We very much appreciate your insights.

Thank you so much for having me. Really appreciate it, Jordan.

All right. And hey, there was a lot there. Don't worry if you missed a little bit. Go to your everydayai.com Sign up for that free daily newsletter will not only be hey, a lot more AI news, but just more insights and more depth into what Sandeep was talking about. Thank you for joining us and we hope to see you back for another edition of Everyday AI. Thanks, y'. All.

And that's a wrap for today's edition of Everyday AI. Thanks for joining us. If you enjoyed this episode, please subscribe and leave us a rating. It helps keep us going for a little more AI magic. Visit your everydayai.com and sign up to our daily newsletter so you don't get left behind. Go break some barriers and we'll see you next time.