AI Virtual Conference: Industry Views on the State of Artificial Intelligence
Earlier in November, FINRA hosted the virtual AI Conference to bring together regulators and leaders across the financial services industry to discuss the use of artificial intelligence and related opportunities and challenges.
On this episode, we’re dropping in for a quick listen the first of the conference’s sessions, “Industry Views on the Current and Future State of Artificial Intelligence,” with Imperative Execution CEO Roman Ginis, SIFMA Managing Director and Associate General Counsel Melissa MacGregor and Head of Fidelity Labs Mona Vernon, moderated by Haime Workie from FINRA Office of Financial Innovation, as they talk about the role of AI in the securities industry.
Resources mentioned in this episode:
White Paper: Artificial Intelligence in the Securities Industry
Listen and subscribe to our podcast on Apple Podcasts, Google Play, Spotify or where ever you listen to your podcasts. Below is a transcript of the episode. Transcripts are generated using a combination of speech recognition software and human editors and may contain errors. Please check the corresponding audio before quoting in print.
00:01 - 00:22
Kaitlyn Kiernan: Earlier in November, FINRA hosted the virtual A.I. Conference to bring together regulators and industry leaders across the financial services industry to discuss the use of A.I. And related opportunities and challenges. On this episode, we take an abridged look at the first session of the conference on industry views on the current and future state of A.I.
00:22 – 00:32
00:33 - 00:49
Haime Workie: I'm Haime Workie and I head up FINRA's of Office of Financial Innovation. Today's first panel will be looking at industry views on the current and future state of A.I. And what it means for the securities industry. So why don't we kick things off by having each of our panelists briefly introduce themselves? Roman, let's start off with you.
00:50 - 01:04
Roman Ginis: Hi, I'm Roman Ginis, the CEO at Imperative Execution. We brought the InteligentCross ATS to market two years ago and it was the first equities ATS to use A.I. as part of the matching process to minimize market impact.
01:05 - 01:07
Haime Workie: Thanks, Roman. Melissa, you're up next.
01:09 - 01:23
Melissa MacGregor: Hi, my name's Melissa MacGregor. I'm a Managing Director and Associate General Counsel at SIFMA, where I've been for 15 years working on technology and regulation issues, including privacy, electronic record keeping, outsourcing and more.
01:23 - 01:25
Haime Workie: Last but not least, we have Mona.
01:26 - 01:48
Mona Vernon: Hello, everyone. My name is Mona Vernon. I am running an organization called FinTech Fidelity Labs, which is a fintech incubator within Fidelity's investment organization. And my other hat, which fits the topic a little bit as I sit on the board of a nonprofit called Fintech Sandbox. And I'm delighted to have this conversation with you.
01:48 - 02:07
Haime Workie: Well, thank you to all of our panelists for joining us today. When we first start off by thinking about what is and what it means, we've heard the term artificial intelligence used to describe a number of different techniques. And in the FINRA white paper that we put out a few months ago, we tried to address this issue. Mona, how do you go about thinking about A.I. and what it means?
02:08 - 02:11
Mona Vernon: I'm not going to spend too much time on the tools. There's the gamut of them from supervised to unsupervised learning, but let's start with what I think they can be doing and be valuable for firms in the short-term and the long-term. And it's really about when, first and foremost, as a precondition is the investment in time required to extract or create the data that feeds those tools in this model so that you can glean impactful insights, so that's one element.
The data is incredibly important. And in order to do that, there is a set of work that needs to happen to galvanize an organization to make the data useful so that you can apply those tools to extract the insights. And so the example of that is, there is not only structured data within your organization, but there is also many types of unstructured format of data sets like PDFs, PowerPoints, blogs, even voice or images that constitute data sets that can really be analyzed with more sophisticated tools in that AI bucket.
And then there is a component that I think is interesting to highlight, which is those tools are really valuable when you get to the last mile of adoption, and so then it really leads to evidence-based decision-making from the output of the analysis. So in the short-term, I'll finish by saying that I see a lot of value in actually using these type of tool sets to extract, transform and start sorting the data in a more significant way to then feed more decision support tools leveraging the same AI tools, and in the long run, I continue to see meaningful insights that a firm can extract from all this analysis of these vast sets of data.
03:50 - 03:51
Haime Workie: Mona, I like that phrase, the last mile of adoption. Melissa, I'll turn it over to you. It's the same question but your thoughts about kind of what AI means and what types of tools are being deployed in the securities industry.
04:02 - 04:02
Melissa MacGregor: I think right now, there's a lot of different definitions of AI, particularly in the securities industry. I think even FINRA's own paper included three different definitions of AI, and although I don't think it's necessarily the absolute necessity to have a common definition, I do think that there should be some sort of umbrella definition that focuses on the characteristics of AI. So looking at artificial intelligence systems that act in the physical or digital world by perceiving their environment through data acquisition, interpreting the collective data, reasoning on the knowledge or processing the information derived from this data and identifying the best actions to take to achieve the given goal, AI systems should be able to adapt themselves to their own algorithms by analyzing how the environment is affected by previous actions, knowledge or data.
And I think we're seeing SIFMA members using AI in a lot of different ways, particularly in some trading functions, surveillance, AML and a KYC analysis, some financial operations and then also in cybersecurity, so I think we'll see expanding uses of AI as the applications develop further and as we begin to better understand what the most effective uses are.
05:24 - 05:25
Haime Workie: Thanks, Melissa. What about you, Roman? Your firm uses AI techniques within their ecosystem. How do you go about thinking about what AI is?
05:38 - 05:39
Roman Ginis: So, one common definition of AI is called the process definition. You can think of it as a system. An AI system is one that measures something in the environment then learns from the observation that it just saw as well as historic operations and then can act on the environment in order to achieve some kind of objective. A good example of this would be if you're building an automated car that tries to stay in the middle of the road, for example, with an advanced cruise control, and the instruction from the driver is to stay in the middle of the lane and to stay a certain distance away from the vehicle in front of you.
So, an AI system that is widely used today in many vehicles will have a radar or sonar that measures your distance to the vehicle in front. It knows your speed. It can also use cameras to measure the distance to the sides of the road, and so every measurement is an observation which tells the system how close it is to the vehicle and whether it should do something about it. So you have a series of measurements which then get incorporated as a data stream into a machine learning function, which observes something in real time as well as has been trained on many miles of road driven and then can act by turning wheels left or right or speeding up, speeding down depending on what needs to happen. So, this is a process definition.
Another definition that's widely used is one that refers to how close a system is to what a human would do in a given situation. So, for example, if you're evaluating a data set, and you're looking for a solution, if you can build a machine that evaluated the same data set, comes up with the solution that the human would, given enough time and energy invested, then that would be an AI system. So, these two definitions have been around for many decades.
I've been in the business of building smart AI optimization systems for most of my career, and the system that we use in IntelligentCross ATS for matching buyers and sellers really falls nicely within both of these definitions. We match orders with the objective of minimizing post-trade price stability so that the matches result in trades that happen as close to fair value as possible, so that's a very clear mathematical objective. You want the price to stay stable. You can measure your deviation from stability after each trade. So, we measure the price movement after each trade. We incorporate it into a machine learning function, which has been trained on all the trades that the ATS has done in the 2 years since we launched it, and it can act by adjusting the timing of the subsequent matches in the system to stay closer to fair value.
The second criteria for AI systems that I just talked about, the similarities to what the human would do, and so in this sense, if you look at the trading history of IntelligentCross, and you wanted to understand why it matched at a certain time, you can clearly see from the time that we see market data events, changes in prices as well as the trades that the system has done, based on the objective of getting close to fair value, you could reasonably say that the human would have arrived at the same decisions. So that's how we think about this.
09:14 - 09:15
Haime Workie: That was really helpful, Roman. And how have you seen the evolution of AI over the last last few years, and what do you think it's likely to do kind of going forward?
09:23 - 09:24
Roman Ginis: So, in 2006, we witnessed probably the most important event in computer science, and that was the invention of what AI researchers called deep learning. Professor Hinton at the University of Toronto has discovered a particularly fast and effective way of training neural networks that is vastly better than any AI process that we have built in the past.
So neural networks have been around since 1950s, but they've been incredibly difficult to really scale and use. And what Hinton has discovered in 2006 is that you can train neural networks on a conventional graphics processor because graphics processors that all modern computers have are particularly good at matrix manipulations, and you can train neural networks tens of thousands faster on these pieces of hardware that is basically available in your home PC.
And the results were so dramatically better than anything that we have designed in the prior decades, it completely changed pretty much all the training processes we have today. So, the reason why we have cars that can stay within lanes and Teslas, and many manufacturers are working on the fully automated cars, all of that is due to the invention of deep learning and so the applications in video processing, in voice recognition. This is why you have Siri and Google Voice, all came out of this new innovation, so this is probably the most exciting thing that has happened, and going forward, there are several thousand research papers coming out and various applications of deep learning in every sector including the financial sector that are really exciting, and at the same time, there is a range of techniques, some of them are more applicable to video processing and maybe less practical to financial services, and I can speak more to why our industry is materially different from the way these systems are used in video and voice.
11:28 - 11:29
Haime Workie: Thank you, Roman. Melissa, Mona referenced this idea, of the last mile and how AI systems can be used to assist people in making decisions. What do you think about this issue? Are AI systems today being used to make decisions themselves, or they're more of a tool to help actual human beings make decisions?
11:47 - 11:48
Melissa MacGregor: At this point, AI systems are not generally being used to make critical decisions for our member firms. There are AI functions that are being used to help employees at our firms to make better decisions and to perhaps remove some of the necessary routine functions that firms are having to deal with, that perhaps this might be a place where AI could help them be more efficient and to perhaps respond to investor requests that are more routine, thus allowing firms to deploy the human resources in ways that are more effective and efficient and can honestly be more helpful for their clients. We probably will see more advances in this area, and we'll likely see more decision-making, but in the end, humans are still necessary.
12:42 - 12:46
Haime Workie: So, I almost think of this as the bionic man concept. So, you have an automated system that kind of assists with the human's capabilities. What ways do you think this will potentially change or impact firms' business models, their processes or their practices?
12:57 - 12:57
Melissa MacGregor: I think a lot of firm practices will only be enhanced by this. I think we have seen a lot of new business models, particularly in the broker/dealer industry over the past several years, but in the end, I think that AI is not the main driver of those new business models but perhaps is helping to enhance them or push them forward a little more rapidly than we would have expected. But at the same time, I think AI will only enhance what our firms are doing from compliance and business trading, et cetera, everything that they do. And who knows what the next technology will be? Just like when the Internet took the world by storm in the '90s, our business models may have adapted to that, but in the end, our firms are still functioning in the same way, just using different platforms.
13:44 - 14:39
Haime Workie: So, let's switch gears a little bit and focus more on the specific areas that we're seeing AI being used in securities industry, and let's start off with the investment process. So that can include a number of different things. But why don't we think about trading first, Roman, you describe one example of how AI can be used in the context of assisting in the execution of trades. We know that AI is used in a number of other areas of trading. So in the front end, for example, being used to help track data after data has tracked down, for example, social media sentiment data or other types of alternative data, being able to process that and being used as information flow to help gain Alpha, as well as potentially assisting with portfolio management. As AI is being used in all these various types of investment processes. Melissa, do you see any associated risks either specific to trading or otherwise and how are firms thinking about these risks and seeking ways to mitigate them?
14:40 - 14:40
Melissa MacGregor: I think firms are working really hard to try to mitigate any risks associated with AI. Obviously, the time it takes to implement an artificial intelligence system is generally very long, a year or longer because those systems need to learn before they can be put into use. So, having those data inputs is critically important. In the market trading systems, I think there's some issues around volatility and unexpected market impacts, may they be elections or storms or whatever they might be that week. The systems may not always be able to account for those changes, as they come rather quickly and cannot be relied on to look at history, as some of those of events may never have happened in the past, so that's always a challenge.
I think there's also some issues FINRA's report noted that systems could learn from each other, causing collusion and other issues. Obviously, that's a scary concept and makes us think of all of the scary robot movies that we've seen in the past about computers taking over the world, but I don't think we're quite there, and I think that firms will be really in tune to those challenges and try to head them off wherever possible.
15:54 - 16:10
Haime Workie: Thanks, Melissa. So, in addition to the investment process, another kind of big bucket area of where AI is being used is communications with the customers and enhancing that customer experience. Mona, can you tell us how firms are currently using AI to enhance the customer experience?
16:12 - 16:13
Mona Vernon: As you may have experienced it, hopefully in a positive way, but sometimes it's a little challenging still, is there is definitely intelligent chat box and virtual assistance have been around for a while, and I think they're getting better and smoother, and they're delivering. And then there is application and customer identification through voice biometrics, for instance, and supporting a more personalized customer experience is part of what this kind of tool kit can help deliver.
I think what's really important in this moment to talk about is that these tools are also instrumental in this COVID-19 crisis in helping accelerate the digital interaction with clients and customers and communicating digitally and this push in eliminating paper. So one example, as Roman described, is there has been some significant advances in AI that is actually leapfrogging a lot of the natural language processing tools to a next level, and so I'm pretty confident that we're going to see an increased use of some of this combination of using OCR and NLP to extract information from a form--
17:18 - 17:21
Haime Workie: Explain to our folks who may not know what OCR is.
17:21 - 17:22
Mona Vernon: Right.
17:22 - 17:23
Haime Workie: Maybe a little bit of layman's terms.
17:23 - 17:24
Mona Vernon: Sure, so OCR is the technologies used for recognizing text in a scan, so if you take a document and put it in your scanner, and then that is digitized not just in an image, but you actually get the words and numbers in a way that's usable for the downstream software, and in this case, I would feed it into a natural language processing set of algorithm. And so that's the ability of having your existing documents turn into something that's digital and usable, something like being able to no longer fill the same form several times, for instance.
It's this ability to take things that were done in paper with existing paper and turn it and have it exist in a digital format so that it's not only easy to access and doesn't require that friction between the customer and the firm, but also so that you can run this type of AI tools on top of this data set because it is formatted in a way that it can be fed as datum into the models.
18:25 - 18:26
Haime Workie: What do you see as being the way that AI may be used by customers in the future, or firms to interact with customers in the future?
18:33 - 18:33
Mona Vernon: I'll pick up on that bionic theme you brought in. I think it's really important to think about the design of the experience and the workflow so that it removes friction and it facilitates a better experience for the customer. At the same time, there is experiences that do require that human interaction. We think about having the machine have their place, and it's not a substitute to replace a human contact.
When I think about someone who's connecting with us during an experience that's a life event like the loss of a loved one, it's certainly an experience where you do need the human touch, and so I think it's about finding places where the automation makes sense and enabling that human connection to be done with more efficiency, and perhaps even with more empathy because of the access of all that background information about the customer.
19:29 - 19:31
Haime Workie: Thanks, Mona. And Melissa, how do you see AI impacting the role of firms, and specifically how the firms may interact with their customers?
19:40 - 19:40
Melissa MacGregor: Yeah, I think there's a lot of really interesting developments coming in this area, particularly new tools for financial advisors, and picking upon on what Mona just said about a death of a loved one, I think that there's a lot of interesting tools out there that firms are starting to use or using in a lot of depth now where these applications can review all the communications from that customer, perhaps various social media posts or whatever information that the firm has available about a customer, and can synthesize that information and perhaps cue an advisor or someone else in the firm's front office to perhaps reach out to a customer who's maybe experienced a significant life event, such as the birth of a child or a death of a loved one, that may impact their financial situation and may cue the advisor to perhaps reach out and offer assistance in any way they can. And I think there's a lot of really interesting and hopefully helpful ways that our firms can interact with customers with the of assistance of AI.
But as Mona said, human touch is still extremely important for many customers, but not necessarily for all customers, and I think, as time goes on, we're going to see less interest from some customers in interacting with a human, and I think we've seen some business models which thrive on that and that they would, generally, younger investors want to just access the firm, look at their applications and perhaps a chat bot, but don't necessarily need or want to have phone calls or other direct contact with the firm. I think we'll see an evolution of that as the population ages and perhaps see some additional changes as time goes on.
21:30 - 21:30
Haime Workie: So, another category or space where we're seeing AI used is in the context of operational functions of firms. So, what I mean by this is things like compliance, risk management, administrative functions. Mona, with respect to these types of functions, what are some of the applications that broker-dealers are exploring?
21:48 - 21:49
Mona Vernon: That's a good question. So, I think there is several areas in compliance and risk management where AI can play a significant role. The essential factor for success, however, and I'm going to repeat that, is access to validated and robust data sets. The integrity of the data and the governance structure of the data is paramount importance. And, that being said, the reason I made that point is that a lot of the data is proprietary, and so companies must invest in curating and preparing the data. That's as many of us that have been in that AI space would say, it's about 70 or 80 percent of the work.
In the compliance space, I'll go back to your question, we've seen a proliferation of use cases encompassing marketing content, public communication analysis under FINRA 2210 to due diligence analysis and adverse media negative news analysis to accurate identity verification under AML and FINRA 3310 and KYC as part of the BSA, Bank Secrecy Act, so I think this is actually a space that's going to continue to really benefit from the maturation of those tools. As Roman was pointing out, the significant advances that are starting to really become more accessible and mainstream, and as the data creation is becoming more robust. So, and more broadly, outside of that specific set of compliance and risk management, there's actually interesting and exciting application in the cybersecurity risk space where patterns of malicious behavior are hidden in the set of binary [audio drop], malware code, and only a machine can quickly identify those permutation of complicated and obscure patterns and exploits those risk.
And then even in places that are what I would call more traditional financial risk management, we've seen credit risk analysis applications that are both for rated and nonrated counterparties to get a better sense of economic strength, and for instance, in that space, logistic regressions analysis have been used in credit analysis for decades to predict the probability of a default.
And I would say there is some evidence that models using things like gradient boosting, to name just one method, but there are many others, can outperform logistic regression models because of the ability to capture more complex relationships in the data. And so I think that's another area where we've seen a lot of FinTech innovation, and I think the maturation of this tools and technologies around AI with a more interesting data sets is going to continue to be an area of development, and I think it's a worthwhile endeavor.
24:25 - 24:26
Haime Workie: So, Melissa, Mona went over with us, very skillfully, a number of different areas where AI can be used in the context of a firm's operations. How do firms go about exploring what are the uses cases and finding out which area it belongs and which area it doesn't belong?
24:42 - 24:42
Melissa MacGregor: I think firms assess AI like they do any other choice they make in their business. I mean they obviously look at need, and to the extent that there is a need, then they will look more closely at implementing the new technology. So, one thing that's driving AI in a lot of business functions is the need to analyze large quantities of data. We all know that big data is becoming an ever-increasing member of our firms' decision-making processes, so looking at how firms need to process data obviously will drive the need for certain technology, including artificial intelligence.
I think then they also look at effectiveness, how effective can the technology be to assist in the business or the surveillance or whatever other purpose they might be looking to use it. Cost, of course, is always a factor, and if the benefit doesn't outweigh the cost, then perhaps they'll take a closer look. Costs, obviously, are coming down on a lot of this technology, just as always seems to happen with technology, that the cost over time does seem to decrease as it becomes more prevalent and there's more competition in the space.
So, there's areas where AI has become more important and more prevalent than others. I think in compliance we're seeing a big increase in AML compliance, where AI allows firms to analyze thousands or millions of accounts and transactions over a long period of time to assess for perhaps any number of financial crimes that might be being committed across those accounts and transactions, which would not be possible with human or it'd be much more challenging, I should say, with only human eyes on all those transactions and more simple computer calculations, whereas artificial intelligence can look at trends and how transactions are made and timing and so forth to better assess when those transactions might need to be flagged for additional human review. So, I think we're going to see that increase in various areas in compliance.
Particularly where we're seeing some firms implementing AI in electronic communications review, again, for the same reasons. Looking across multiple channels for electronic communications is challenging for an individual reviewer, whereas an artificial intelligence system can look at those communications across multiple channels and perhaps find troubling contact or perhaps increased risk where the firm may need to take a closer look at those communications. I think there's a lot of ways that firms are looking at those new technologies and deciding where to implement them, but in the end it's just like any other business decision when you're looking at cost and effectiveness and the need for that new technology.
27:33 - 27:34
Juan Echeverri: We do have two questions. The first one for Roman, "Within the framework of artificial intelligence, could you comment on the ability to create an audit trail that reflects the changes and decisions being made to the firm systems?"
27:50 - 27:51
Roman Ginis: Certainly. So being able to have a system that is introspectable, where you can analyze each decision made and what caused the system to make the decision is obviously crucial in our industry. So intelligence across ATS is regulated by the SEC and FINRA, and like any ATS, we preserve the history of all orders received, the trades, but in our case, because it's an AI-enabled system, we also have a history of the recommendations that our AI system has made about wanting to match orders.
The process of training our system that involves the data we have collected and how we decide when to schedule matches is also part of our record. So, in general, just to broad the question a little bit, as we build AI systems for the financial industry, it is critical to think about what kind of introspection ability we want to have looking back at the decisions that they've recommended making. One of the big concerns is our AI stuff is black boxes. While a complex system can have many, many weights, you may not be able to say why a particular weight is 4 1/2 instead of 5, I think what is possible, and something we can reasonably expect of these systems, is being able to say, "Well, these were the inputs that we trained the system on. This was the decision that we would have liked the system to have made given those inputs, and what did the system actually do?"
And if the system tracks reasonably well, what we expect the system to do, which we can measure, let's say, with 80 percent probability or 90 percent probability, then we can say, "Well, we've designed and trained the system to do what we, as industry professionals prefer, and we can think of any deviation from that target as opportunities to reacclimatize and improve the system."
29:51 - 29:52
Haime Workie: Mona, I will turn another question over to you. Can you tell us about some uses or enhanced opportunities for individuals to participate as investors and to promote financial inclusiveness as a result of AI?
30:05 - 30:05
Mona Vernon: Yes, thank you for asking that question. The ethical use of AI can be more participatory, and AI's promise that it will deliver real business value by helping firms be more efficient at finding and retaining customers. We do that by looking at complex relationship in the data without any preconceived biases.
However, if the data embeds historical biases and prejudices against certain demographic groups, the results will be biased, and to combat this, firms must ensure that they practice participatory AI, which means that the data-governance structure must accommodate for when we remove historical bias that we know exists in many financial and social data set. All of this means having people at the table who can share different experiences, and this inclusivity paradigm is a chance to build a better economic system that benefits all.
31:02 - 31:03
Haime Workie: Thanks, Mona. Well, thank you to all of our panelists for sharing your views with us today.
31:08 - 31:23
Kaitlyn Kiernan: That's it for this episode of FINRA Unscripted. We're going to take a break for the holiday season, but we will be back in the new year with new episodes. May you and your family stay safe and sane through the end of 2020. Until next time.
31:23 – 31:28
31:29 - 31:56
Disclaimer: Please note FINRA podcasts are the sole property of FINRA and the information provided is for informational and educational purposes only. The content of the podcast does not constitute any rule amendment or interpretation to such rules. Compliance with any recommended conduct presented does not mean that a firm or person has complied with the full extent of their obligations under FINRA rules, the rules of any other SRO or securities laws. This podcast is provided as is. FINRA and its affiliates are not responsible for any human or mechanical errors or omissions. Parties may not reproduce these podcasts in any form without the express written consent of FINRA.
31:56 – 32:02
Music Fades Out