 Can you hear me okay? Good. Good. I'm gonna walk around. I've been in three continents in four days, so if I stand still, I'll probably drop over sleep at some point. Thank you for the invitation to come here. I'm very happy to be here. I'm very excited to talk about one aspect of Fintech. Fintech is having a lot of effects on financial markets and the provision of financial services, and I think in no area is it having more concrete effects already than an artificial intelligence. So I'm very excited to talk about this. Now I think I have about 20 to 25 minutes to talk and I'm gonna leave about five minutes for questions, hopefully. I get excited and I'll give you lots of stories so maybe I'll eat into some of that time. I have to start with a disclaimer. I'm speaking on my own behalf, not on behalf of the governors of the Federal Reserve of the Federal Reserve System, so please don't get me in trouble by saying I'm speaking on their behalf. I wanna talk about three things. First, I wanna try to give you a definition of what artificial intelligence is. I think with 100 people in the room, there's probably a hundred different ideas of what it is, and maybe they're not all the same. So I wanna try to sort of level set so that we're all on the same page and try to give you a sense of the difference between artificial intelligence and big data and machine learning. I wanna talk about how it's being used in finance and I think this is really important because there are so many different use cases and they are across a wide spectrum of parts of finance. And I think this is really important because when you go to a conference on blockchain or distributed ledger, you'll hear a lot about the proofs of concept. Here's all the things we might be able to do with this technology at some point in the future. With artificial intelligence, every example I'm gonna give you is something that's already being utilized, already being offered by a company for financial services. And then as I talk about these use cases, I think you'll start to see the financial stability implications as I try to sort of describe different ones and at the end I'll use a couple of slides just to bring them all together. So what is artificial intelligence? As I said, you probably all have some sort of concept in mind but if I gave you a one question test and said write down a definition, you'd probably struggle to exactly put that into words. So I do wanna start by trying to sort of knock down some of the misconceptions about what artificial intelligence is. Okay, it's not the scary monster from the Terminator movies, okay, this thinking machine. It's not scary, it's not science fiction either. It's real, it's being used now and nothing like this has happened yet. At least, hopefully it's not even on the horizon. The opposite end of the spectrum are people who think it's more like this magical thing. Well, you get all this data and something spits out this number and I don't really understand what it comes from and it's not, you can't possibly understand where it came from. That's not what artificial intelligence is either. I think the best picture that I've seen for what it is to try to understand it is this one which is taken from a study done on diabetes. This looked at about 10,000 patients, looked at over a hundred different pieces of data for each of those patients and did some clustering analysis, one type of artificial intelligence to try to identify are there specific types of diabetes and what you see here is very clear representation that there's three types of diabetes. Now, this was actually a very landmark study because for a long time, they had assumed there were two types. Now, there were some people, some doctors who had suggested that, yeah, I think there might be another type that we haven't quite identified yet and even though the human beings were able to sort of guess that there might be one, it was using these data techniques to identify the third cluster that sort of captures the power of this and this was published in, I believe it's a journal of nature, so very high end journal. Now, another picture that sort of gives you a sense is something like this. It is mathematical and it can be complex math, it can be a combination of many different types of math or it can be simple math and just layered upon itself so that you get much more robust structures. So when you think about artificial intelligence, I hope you think of something like this or that picture that I just showed you. Now, I need to give you some definitions, okay? I'm gonna start with big data because this is the simplest one. I'm gonna give you three definitions. All of them are taken from an FSB report and if you get access to the slides, if they get posted on the website, the final slide actually has a link to it and the citation. Big data is just the storage and analysis of large or complex data sets, okay? This is becoming more and more common because storage is so much cheaper so we can get more data which means there's a greater supply of people wanting to sell us this data and now we have the processing power to analyze this data. So this is not new but it's becoming more and more common. Now, this is the important one. What is artificial intelligence? So if I gave you the one question test, this is the answer that I'd want you to give me, okay? The development of systems, able to perform tasks that have traditionally required human intelligence and I think that's the key there, traditionally required human intelligence. So the most common example, the easiest example to think of is language processing. 10 years ago, if you spoke to your computer, it sat there and did nothing or if you spoke to a little device sitting on your laptop or in your smartphone, nothing happened. But now you can talk to them. They process language. They're using artificial intelligence to process language and there are even applications out there that will now translate what you're saying into another language in just about real time. So it's almost like out of Star Trek, okay? But there's other aspects to this as well. Humans traditionally were the ones that identified patterns. They said, I think there's a linear relationship in this data and they ran some sort of regression and they found a result. Now we can set the computers to look at these data sets and say, we think there's a pattern. Go see if you can find it. Even if we don't tell them what the pattern is that we think might be there, this is what we mean by artificial intelligence. Machine learning I think often gets lumped in with artificial intelligence and I do think that's somewhat appropriate because I think it's a subset of artificial intelligence and this deals with the specific algorithms that are supposed to optimize things, okay? Using whatever experience they've had on previous data sets but the idea is that it's done with little or no human intervention. In some cases you give it a whole bunch of data sets and say, these are the outcomes that I like from these past data sets replicated on these data sets, okay? Graphically I took this again from this FSB report. If you start with artificial intelligence, you can do that with big data sets or without. So there is some overlap between them but they don't have to be completely connected. So I think we can separate big data and artificial intelligence at times. And the same is true for machine learning. I think it's a subset of artificial intelligence and you can apply it to big data or you can apply it to traditional and smaller data sets. So I hopefully have help level set here. So how is it being used in finance, okay? I wanna start with the consumer facing applications because these are the ones that we're all aware of, okay? I have some great stories about chatbots and virtual assistants but they have very limited connections to financial stability so if you wanna know ask me at lunch. It's hard for me to bite my tongue because they're some of my favorite stories. Insurance pricing. There's some really interesting things going on out there. There's some Chinese websites, merchant retail websites where you can go and make a purchase and when you make that purchase, you click purchase and the next screen it says, would you like to buy insurance? For example, if I wanted to buy this tie I wanna make sure I got it before I left for this trip. It might have said, do you wanna buy insurance to make sure this shows up on time? And it shows you the price right there, 17 cents, something like that. Now that price is actually determined for each individual item after you click purchase. It looks at where the retailer's located, where you're located, looks at a whole bunch of similar transactions, figures out how often it was late and comes up with a price. That's done in nanoseconds and they do that using clustering techniques and artificial intelligence. Fascinating application for this micro insurance market. Credit scoring is one that's gotten a lot of attention. We've heard a little bit about fintech credit today. A lot of these fintech credit providers are assessing people without a traditional credit score and they do that using non-traditional data. They might look at social media accounts. Some of them look at things like Amazon purchases and they do this and they're able to assess credit in a different way. That might have a financial stability implication. It could have multiple. One is that it increases potentially financial inclusion. People who didn't have access to credit now have access to credit, for better or for worse. The other thing might be that these are non-traditional, non-tested methods of assessing credit and perhaps increasing credit to the wrong people at the wrong time. Now some of the things that we don't hear as much about are the operations in the back office sort of issues and where this is being used. A lot of this has been going on for years and we just haven't kept up with it. Some of the most interesting I think are the capital optimization uses. So there are firms out there that are basically saying, look, you hold too much capital that's costly for you. You hold too little capital, you're gonna get in trouble. So you wanna hold just the right amount of capital. But it's a very complex decision and every day asset prices are going up and down. Your own strategy is changing what your balance sheet looks like. There's just too many variables to do this in a very sophisticated and very efficient way. But they develop tools that will help firms keep the best amount of capital at any given time, forecast when they're gonna have to purchase more, let them know that you won't need as much capital in the future, cancel those capital raising plans. Very similar stress testing. Firms are stress testing themselves. We're stress testing firms. And often, the results from year to year are gonna swing wildly. So firms wanna know what are the key variables? What are the things that I need to keep my eye on and that you're using now artificial intelligence and machine learning to figure out what those key variables are. Trading and portfolio management. This is sort of where you really start to get into some of the applications that have raised some eyebrows and made people say, maybe this isn't good. You've got this sort of proverbial black box just spitting out things. Trade execution. If I wanna place a big trade and I don't wanna have an effect on the market, I need to know where I should place it amongst all these platforms out there. Which one should I give? How much? When should I give it? How long should I wait in between slices to give this out? You can certainly have a human being sort of figure this thing out. But when you're trying to do this in nanoseconds, you need the machine to figure this out. Identification of price signals. I heard a really interesting application here. It had to do with commodities futures. And for a long time, commodities futures, something like wheat or grains, you wanna know about weather. So you've forecast, you get forecasts of weather from national weather services or international organizations that are forecasting weather and you feed them into your models. But now what they're doing is taking live feeds from weather sensors. And these weather sensors will tell you that in the top 10 areas, it's currently raining, here's what the sunlight is. And so in real time, they're updating their forecast for crop yields based on real-time weather. So certainly you could have a human being doing this at the computer, having processed in this, but it's better to do it with a machine that can take this exact same data and process it in real time. Fascinating application. Compliance. We're starting to get a lot more into the financial stability implications here. Firms out there nowadays have a lot of things in which they have to comply. New regulations, old regulations, things that they've been doing that they never had to think they had to monitor and now they're trying to monitor it. Monitoring behavior is an interesting one. After the LIBOR scandal of five, six, seven years ago, maybe eight years ago now, a lot of firms are trying to monitor the behavior of their employees more actively. And there's a firm out there that does this by monitoring text messages, phone, email, everything in real time. And they build this system where it brings all the information in and analyzes it in real time. And I saw a demonstration of this a few months ago. It was really interesting because they took some of the actual LIBOR conversations that took place via email and they ran them through sort of in real time and they show you these lines coming up on the screen and it was these two traders, the one demonstration I saw and they were talking about a birthday present. The one trader was saying, I'd really like a birthday present, my birthday's coming up. And as these lines scroll up on the screen, you see just sort of a green light. Then three or four lines down it starts flashing orange and then it flashes red and it identifies suspicious behavior. And after like 10 lines, it says possible market manipulation. It was just fascinating how quickly it was able to sort of analyze this data. So monitoring behavior is one area where we might be able to see uses of this and see uses which could help us to avoid big problems that we've had in the past. Another area is data quality assurance. We have a lot of regulatory requirements where people are supposed to fill out forms, give us more data. Obviously they want to avoid the types of errors that are going to lead to embarrassing mistakes, whether it's fat-fingered errors or sending in the wrong data. Traditionally we've had human beings looking at this and somebody who's looked at it for 20 years looks at it and says, well, that's clearly a bad number and they just sort of develop an understanding for what's good and what's bad. But now they have programs that analyze that same data sheet and they look not only for fat-fingered errors, the obvious, wow, that's actually supposed to be a million, not a billion, those sorts of errors. But they also look at trends. Well, this data was trending up and now it's trending down. Are you sure that data's correct? They also look at correlation, if you submit 1,000 variables, they will look at the correlations of all those variables and identify that traditionally all of these 73 variables tend to move in the same direction and three of them didn't move in that direction this time. This is something that a human being is not gonna capture, okay? But now these programs can do it in a much more active, much quicker, much more real-time basis. The flip side of this is the supervisors, okay? We are now trying to do all the compliance monitoring for these firms, okay? If we get that same regulatory report, we're gonna wanna make sure that the data are accurate. We're gonna wanna make sure that the trends that we see, the trends that we're concerned about are being monitored. So we might not be looking for the same things. I might be looking at the leverage of the firms or the maturity transformation that's taking place, other things than what the firms are. But I could use this sort of technique to monitor that in real time. Fraud detection, the credit card companies are becoming much, much better at identifying fraud, okay? They can do it in real time. They're looking at a much broader set of data, not just where the transaction took place and the amount. What was the good being purchased? How long has it been since you last purchased this good? That sort of thing. Which will hopefully make it easier for folks to travel around the world and not have their credit card stop when they try to make a purchase at a Chinese pharmacy. Not that that happened to me any time recently. Okay, I wanna talk a little bit about systemic risk identification. This is probably obviously of great interest to the European systemic risk board. So let me talk a little bit here. I have a picture here taken from a machine, an artificial intelligence firm used with permission and it'll become obvious on the next slide who this firm is. Typically, we look at patterns like this. The one on the left is the one that economists have looked at for years and they said, oh, look, there's a pattern there, there's a line, I'm gonna run a regression, I'm gonna get a result. I think most of us, if we saw these other patterns, we'd go, yeah, well, I don't know, my regression techniques aren't gonna work on this. I can run a regression on that loop there but it's just gonna give me a flat line because we don't typically use these sorts of patterns. Now, you look at them and you say, well, if I saw this pattern, I would obviously not run a linear regression, I recognize that this is there but this is obviously a very gross oversimplification. We might see these patterns in a hundred dimensions or 50 dimensions depending on what sort of system that we're looking at so they're not always this obvious, okay? Now, using techniques to find patterns like this in data could lead to pictures like this. This is taken from a firm called Ayazdi. They're an artificial intelligence firm, I do not endorse them necessarily, they happen to let me use this picture so I told them I would make sure that their name showed up. What they're showing here, using the same techniques that led to that diabetes picture, they're applying it to about 150 economic variables over 30 years, this is monthly and quarterly data and trying to see if they can identify pieces of the economic cycle. Now, it's certainly not as distinctive as that diabetes picture. You can see some clustering but it's not quite as easy to see. Now, in the presentation and they didn't give me this slide, they then put an overlay on it where it shows the recessions and the expansions, the contractions, all that. It becomes a little bit more interesting to see the patterns that are there. But what they're using are basically those clustering techniques. They're looking at the same kinds of simple patterns, the loops and stuff that we have available to us but we've ignored up till now. Now, this is macro data so maybe not as plentiful as financial data but if you wanna look at systemic risk, you could apply these same things to the perhaps thousands of financial variables that we have and that might allow us to see a little bit more than this sort of picture. Okay, so let me take those use cases and put them into some financial stability implications. So I wanna start by talking about the benefits and I think that's important because when you come from a systemic risk perspective and I work at a financial stability division, we tend to think about the risks and we often forget about the benefits. The risks are a lot more interesting but okay, we'll talk about the benefits first. So easier regulatory compliance. Firms can now use this technology to comply with regulations perhaps in real time. It can save them time, save them money. It makes it easier if they're complying with regulations. If it's cheaper to comply with regulations that should be a good thing for the financial system. Greater efficiency. There could be multiple pieces here. If their capital is allocated more efficiently or they're holding the best amount of capital, that should be better for them in the long run. Or you could think about just the speed with which people could move financing to the right locations. Or it could be something like the financial inclusion angle, okay. If there are parts of the system that do not have financial services because we don't have the right metrics for them and we can now provide them with financial services, this could improve the efficiency of the financial system. We could argue about whether financial inclusion is good or bad for financial stability, but there are arguments in favor of it being a good. Improved systemic risk monitoring, okay. If we expand the group of models that we look at, we can't do worse than we were doing before. We'll have to guard against data mining, but we shouldn't do worse because linear models are just a subset of the models that these techniques are allowing us to look at. If we can improve systemic risk monitoring, then certainly the banks and financial institutions should be able to improve their risk management because they're using these same techniques to improve their own models. Better fraud detection, improved supervisory effectiveness, okay. All of these should be financial stability enhancing. Now these are the ones that I think, like I said, are a little bit more interesting. The elephant in the room when you talk about artificial intelligence and potential risks is the interpretability issue, okay. We're human beings, we like to be able to explain things. When a plane crashes, we want to know why. And ironically we get the black box to figure out why, which is kind of a strange use of the term there. If there's a flash crash in the markets, we want to understand why. So we inherently don't like the fact that these models are difficult, not impossible to interpret. Now I was at a meeting of regulators like yourselves. We had some private sector come in and talk about their use of AI a few years ago and there was a trader, a fixed income trader from a large international bank and he said, every day I'm given a file that contains one terabyte of data. He said I can't possibly use traditional techniques to monitor this data. This meeting took place like a year and a half ago and he'd been using machine learning techniques for three years at that time. So this is four and a half, five years ago now to analyze this data. And he was super excited about the results. He said he made lots of money. So the audience is full of supervisors and so of course somebody raises their hand and says, but what happens if it tells you to do something that you disagree with that goes against your gut or it just doesn't make any sense to you? And his answer was somewhat shocking to the folks in the room. He goes, I don't care. I've made a lot of money using this model and I'm gonna listen to the model. Well you can imagine a room full of supervisors, the hair on the back of their neck bristles up and they're like, oh my goodness, let's get this guy right now, okay? But we should ask ourselves, do we want interpretable models that give you more systemic problems but interpretable or do we want fewer systemic problems but maybe not interpretable? Okay, it's a difficult question. He obviously answered in the one way. I think most of us in the room at the time answered the opposite way. We wanna understand what you're doing. Obviously there's an increased reliance on data here. There's a lot of data privacy, data governance issues but there's also a data dependence issue. If you start using this data, if it becomes a part of your everyday life and then one day the file gets sent with errors in it, that could create a lot of problems. Or if the file doesn't get sent at all, what the heck do you do that day, okay? Presumably you have backup models, things like that. This goes in line with increased third party dependencies. Do we need to look at these third party dependencies, use of these different cloud, AI, big data providers, differently than we would other third party dependencies. There's some school of thought out there that if you use AI it will lead everybody to the same model. Whereas if human beings pick the model we have idiosyncratic biases which cause us to sort of stray from the optimal model in random ways. So we get a selection of models, interesting perspective. And then there's the human resource deficiencies. I don't understand these models. I'm guessing that I heard this morning that you're hiring more data scientists, which is great, okay? I bet you don't have enough yet to keep up with the banks and the insurance companies and the others. My last slide, I just wanted to end with this thought which is that artificial intelligence is not a product or a service in and of itself, okay? It's a tool, it's a foundation which you're gonna use to build products and services. And because it's such a foundational level change in the way we do things there will be innovations on top of that and innovations on top of that. And this could be the start of an innovation spiral. So when people tell me like, oh yes, you know all those things you've said I understand those, I'm not so concerned. I just sort of say that's right. But when innovation spirals occur, you can understand the first generation but the second and the third and the fourth generation these things are gonna change things in ways that you can't imagine. If you think of cryptocurrencies who thought of smart contracts and ICOs and things like that two years ago, okay? Very few people. And with that I will finish and there's the slide with the additional information. And it looks like I have about three minutes to take questions if anybody has any. Please. Beside point estimates we are at least as concerned to get like confidence intervals. Now maybe this is a very naive question but what is the kind of artificial intelligence equivalent of a confidence interval? I mean, if you recognize a pattern how can you be sure especially if these are highly nonlinear that these patterns are kind of quote unquote robust? So I'll answer that quickly because I have to say I'm not a data scientist. I don't pretend to be one. I don't know the answer to the question other than I do know that they're able to do statistical significance tests and I would presume it's very similar to doing a regression. It's just a different set of statistics and calculated in much more perhaps sophisticated ways or maybe it's just as simple as calculating your standard error bands. So I don't know. I've been in audiences where that exact same question was asked of the AI professionals and they said, oh yeah, that's simple. We can do that. So I'll trust them for now. But I don't know the answer. I don't know how it's calculated and because I'm standing in between you and lunch I think everybody just wants to go. Okay, great. Thank you very much.