Skip to content
Transforming Trade Finance Compliance Through Machine Learning and Natural Language Processing episode artwork

Dec 22, 2019 25 min

Transforming Trade Finance Compliance Through Machine Learning and Natural Language Processing

from Risk Analytics Hosted by SAS, GARP.org, Risk.net and others.

At the SAS Global Forum, we caught up with Valeria Sica, Managing Director, Head of Global Trade Services and Jake Jacobson, Technology Advisory Partner, EY about how they are using AI to digitize compliance. (https://www.sas.com/gms/redirect.jsp?detail=GMS116861_161345) Celent recognized Citi as the Model Bank of the Year 2019 award winner for "eight cash management and payments initiatives by Citiís Treasury and Trade Solutions Group which demonstrate how Citi is embracing change and differentiating with digital." (https://www.celent.com/insights/384737214)

Episode details come from the podcast's RSS feed.

Embed this episode

Ready to play

Transforming Trade Finance Compliance Through Machine Learning and Natural Language Processing

0:00 25:32
of matches

Transcript Auto-generated, may contain errors

Hello and welcome. The Global Association of Risk Professionals, also known as GARP, has partnered with SAS to bring you a brand new podcast, The New Age of Risk Analytics, about the evolution of risk and finance integration into the new digital era. Hello and welcome to the New Age of Risk Analytics podcast. We're at the SAS Global Forum.

I'd like to welcome our guests today, Valeria Sica, Global Trade Services Head at Citi, and Jake Jacobson, Technology Advisory Partner, EY. Today at the forum, I spoke on transforming trade compliance through machine learning and natural language processing. We just saw that session. It was a great session.

So we want to share some insights with our listeners who couldn't be with us. So firstly, I would ask each of you to tell us a bit about yourself and the role that you play at the organization you're with. Thanks, Debbie. It's great to be here today.

So I've been with Citi for 26 years, always in trade and in different roles, first in Argentina, and that's where my accent is from. Then I worked for the Latin America region and since 2012 in a global role around trade services. Thanks, Debbie. Thanks for having us here today.

So I have kind of two hats that I wear at EY. I really spend most of my time every day working with clients on financial crimes compliance. I've been doing that for about 20 years through a number of different professional services firms and really everything in and around the people, the process, and the technology changes required to keep up with the evolving set of regulatory expectations and actual threats on the landscape. The other thing I do is as our business evolves, we have to continue to innovate in how we deliver value to our clients, which means that we need to introduce new services, new business models that really deliver outcomes-based business value to our clients.

So I help lead the technology aspects of building out our solutions and managed services business. Great. Thank you. I know you have an exciting project to tell us about because I got the benefit of being in today's session.

But can you tell us what brings you today to the global SAS forum? Absolutely. And this morning we issued a press release as well. So everybody hopefully saw that.

And it's around a project that is designed to detect transactions with potential compliance concerns. So we can spend the time looking at those transactions at data points and trying to see if there's anything that could reflect a money laundering concern or fraud or something that needs to be escalated and taken care of. Do you want to say anything else about the project? No, I think you covered it.

You know, I guess one other element of that is really the interesting and really valuable partnership that EY, SAS and Citi have put together to meet this first-of-a-kind industry challenge. We'll talk more about it. There's a lot of innovation, a lot of really difficult technical and functional problems that we're trying to solve. So the unique value that the three parties bring together really positions us for the best chance of success.

Thank you. I want to step back a little bit. And can you give us some background on what the key objectives are? What are the elements of financial crime that every organization, especially large global institutions, are contending with?

What kinds of financial crimes are we talking about? I think, you know, most of our audience is going to assume we're talking about anti-money laundering, you know, anything that has to do with terrorism financing. Are there other elements of financial crime that this is intended to address, whether it's embezzlement or human trafficking or red flags on any of other issues that I might not have thought of? Yeah, and at least I think you've hit the big topics, right?

But as I say, the devil's in the details. And so as you start to drill in on topics like AML, human trafficking, terrorist financing, the items that we're solving for in addition to the direct AML risk elements are boycott clauses, sanctions, the import and export of military and dual-use goods, certain aspects of fraud that are unique to the trade finance industry. And then there are some other things around the flow of capital in and out of certain jurisdictions that have regulatory requirements around that. It's pretty nuanced.

But really, I think the regulatory driver of this is primarily the AML and the sanctions concern within the trade finance business. Great. Thank you very much. Can you give us some context on how trade compliance has been approached traditionally so that we'll understand what's changing now?

Absolutely. So compliance has been approached in a very manual way. When you think about how trade operates today, like 95% of the transactions are still paper-based. So you get paper, a bill of lading, an invoice, and you need to go through the information that you have in the paper.

That's all you have. And based on that, you have to make a determination on whether there is a violation of any of the compliance concerns that we have. So it's been extremely manual, transaction by transaction. Everything is on a one-on-one basis.

And we rely on very manual processes to know what happened before and trying to understand if this is a transaction that needs to be escalated. In a paper environment, it seems to me since you can't look horizontally, you're going to get more layers of processing. You're more likely to get sort of red flags or hits on them. And does that make the process even more labor-intensive than just the paper?

So that's the challenge we have with the technology. To make sure that while you're using technology in the process, we don't increase the number of false positives. And we've been working and training the engine on the transactions that were escalated over the past three years. So trying to understand the characteristics of those transactions and kind of telling the engine, look for that.

Don't look for everything because if not, you're going to give me a very large number of false positives that will increment the manual process and we want to avoid that. Yeah, I think too, there is a certain inefficiency in the purely manual-based process where operators do lack that time-based and network-based awareness of what's going on. So similar risk items or even identical risk items can be escalated time and time again without that awareness that it's been previously looked at, it's been investigated, we've determined it's not a risk. And so we don't need to keep repeating that process every time something similar comes through.

Okay, great. Thank you. That really helps. What are some of the other limitations of the paper model that new technologies are hoping to address?

So some of the limitations we have in the manual process is to manage large data sets. So we have a lot of information that our operators need to go through and a lot of checks they need to make in every single set of documents to identify the potential concerns. So that's what we are trying to solve for. So we're not asking the machine to make the decision.

We're asking the machine to detect the cases that have more risk. So then we can focus on understanding the interaction of all the data points and making the decision from the current processes and the BAU. So this is more at the level zero or the maker function. It's not to replace the compliance function.

And that to me is very critical at this stage of the development because we are not there yet. So it's more of a detection engine that then can lead to a proper decision. Yeah, I think some of the surrounding technology challenges that we're looking to address is really in the core digitization of the data. You know, 90% of any analytics problem is really prepping and organizing the data so that you can then develop and execute good analytics against it.

And in a complex paper environment, that's really, really a difficult challenge. So, you know, really trying to drive further into deeper and deeper insights extracted from the transactions and the documents involved is going to be critical long term to, you know, kind of the upper bound of potential of what analytics can do in the trade space. And again, everyone thinks that eventually kind of disruptive technologies like blockchain, you know, will really dramatically shift the landscape. But it's going to take some time, given the complexity of the ecosystem around trade and a number of really disparate stakeholders that operate in the space across the entire kind of supply chain.

So, you know, the nature of and the complexity of the documentation within the trade space is rather unique relative to a lot of the other problems we're trying to solve with, you know, these emerging technologies like NLP. I mean, I think one of the things that I keep hearing about these new technologies is that they're not there to replace the humans. They're there to make their systems more efficient and more accessible and to empower them. And I'm getting a message from you that you're finding that's your goal as well, and that it remains, you know, the crux of it is needing the human behind the machine to decipher what you're Yes, because I think we need to walk before we run.

And bringing this technology into our processes today is a big, it requires a big mind shift because it's new. We were always relying on people looking at the transactions on the paper. And now that's going to be replaced by an engine. So we need to understand what the engine is doing and kind of walk by slowly and feel comfortable with it.

Given our industry is very conservative. So, and particularly in this area, we need to be very careful to maintain our standards or even have better standards from where we are today by the added technology. But we want to make sure that we go very slowly and we are comfortable with what the engine is doing and with the hope that it's going to give us more information so we make the proper decisions. That makes a lot of sense.

Now, we've all heard about AI and natural language processing and machine learning and computer vision. Can you illuminate for us what that means in this context? What is important for us to understand about what machine learning is in the trade context and what natural language processing does for this process? Yes, certainly to the point where people are doing what people do best, which is make decisions and apply judgment, as opposed to sifting through paper, compiling information, and really the administrative tasks that the models are much better at.

So what I would like to add is, when we started the project, we gave the team like a few mandates. For example, no rules, absolutely no rules, because rules need to be maintained over time. So we wanted this to be done with the technology in a way that the technology learns by itself over time and how the behavior changes, so we don't have any subjectivity in the project. The other mandate was we need to do this real-time, and that's why the SAS solution really is ideal for this and helping us, because that means you need to manage and process a lot of information in a fraction of a second to give a response because there's a client waiting for that transaction to be processed.

So real-time for us was very important because that's how the process is done today. And we thought that if we are adding technology, it should at least have the same standard or better. So for us, those two points were very critical and really drove most of the reasons to partner with SAS and with MIT to make this happen. One interesting or I think critical element there is as we move away from rules, we move into machine learning and adaptive models.

The concept of explainable AI is really paramount, which is really no black box analytics. It has to be perfectly explainable and documented and understood both by the business who are the ultimate acceptors of the risk, the risk and compliance functions within the bank that are responsible for ensuring proper risk controls, and then the regulators too, to be able to look at a post-factor and say, yeah, we understand what's going on. We're just not ready for, we don't have the track record to be able to implement black box analytics where we just have ultimate trust in the model. Well, I mean, that makes perfect sense.

And it's good that you're, you know, I think there's a perception that regulators are some kind of police officers that come in and shut things down. And what you see in the day-to-day, what we see with the people that we interview and that we talk to is there is an ongoing partnership involved in this, and there is a give and take and that both pushes the other to move forward. We hope it got more part of that process. I was going to ask you a little bit about the explainability.

How do you communicate the complexity of the analytics to people who maybe don't have that kind of programming background? And how are you ramping up the people in those positions who are experiencing this technology for the first time? So as I mentioned in the session, I mean, it's really critical to partner with these functions from the very beginning. So for us also has been an interesting journey and an education to understand what the technology can do, what is real, what is kind of the hype out there that is going to solve everything.

And if you kind of go step-by-step and understanding, and this requires time, it takes time. You need to go through all your policies to understand what needs to be checked, to understand what are the data points and how you combine the policies with the data checks and the engine and kind of evolves at the same pace. So it's not a project to do in isolation and then to show the compliance or we developed this. That will never work in my mind.

That's a personal view. You need to go with them and get their input and kind of design the solution together. So you get to a point where everybody understands more. And EY hosts a lot of lunch and learn sessions in Citi as we go through the project.

And we are learning a lot throughout and kind of getting an understanding of how this works. Yeah, Laura, I'd like to echo your requirement of time because there is certainly a learning curve here, right, for everyone who isn't a data scientist, which is almost everyone involved in the initiative, to understand the difference between feeling comfortable with something and being able to prove mathematically, you know, that feeling. And so, you know, the old standard of take the most conservative approach, make sure that there's absolutely zero risk. That works, but will ultimately, you know, result in replicating the current people-based process within the system.

And that's not really the objective here. The objective is to really change the game in how we're approaching these risks and solving for them and set up, you know, Citi and our other clients in the industry for the next couple of decades on how these risks really need to be managed. Well, speaking of that, how do you foresee the technological advancements here will impact positively a reduction in large-scale financial crime? And what's the kind of arc that is anticipated for this level of partnership?

Well, first of all, I think we're going to get far more transparency into what the true nature of the risk is. You know, we operate on a lot of assumptions around the amount of money that moves, you know, the amount of illicit money that moves through the trade network. And that's based on anecdotal evidence. And the more information that we get structured and organized into data lakes and data models, which will apply the analytics to really understand the nature of the risk.

So that's, I think, point number one. Point number two is, you know, we're going to really benefit by the rapid, rapid advancement of natural language processing and natural language understanding. Because at the root of this problem is getting information out of documents and really a diverse, diverse set of documents. You know, as an industry, we've essentially solved for things like how do you digitize a loan underwriting form?

Right, but that's where the bank fully controls the form, you know exactly which field all the information is in. But if you start to look at the diverse documents that flow through trade, the customs forms, and insurance forms, and bills of lading, and invoices, and across a really diverse set of stakeholders at varying levels of technology adoption. It's not going to work long-term for us to have to map out every document that comes through and tell the machine where to look for each piece of information. As NLP advances and NLP can really make sense of the document and understand the information that's on there, and then bring it into an environment that then we can start to really, you know, take advantage of the analytics.

That's going to be the big, big driver of advancement here, I think. Yeah, it seems to me that the opportunity for patterns that have been unrecognizable, given the silos of people examining paper, obviously the potential for data is that connections will be made that nobody knew were there before. Yeah, if you're a data science professional, you get really excited about challenges like this, right? Because we've got two or three of the hardest kind of AI problems that are out there coexisting within this.

Is there anything else you guys want to tell us about the project or your process at this critical juncture? Because I think there are so many people who are going to end up going through this as a matter of course. And, you know, your insights will surely be welcome to them. So I would make two comments.

One, we are a bank. We are not an investigations unit. So what we are developing is to, for us to comply with our current regulatory requirements. And that's the context and the scope.

So we're not trying to find the third or fourth level of interaction between the parties because our obligation is to manage what is at hand and kind of understand what's going on with these transactions. So to me, that's important with the team to understand really the scope of what we are doing. The second point I'd like to make, anyone who's going to engage in a project like this is, number one, you need to be very patient. It takes time.

You need to be extremely flexible because the goals that you set for the next three months will most likely change as you go through the process and have new learnings on how this can be done. But with all that, we're managing this project within the principle that we are not going to compromise quality for time. So we are not rushing to make this happen because it's very sensitive and it needs to be done right. So we're taking the time that we need to make sure that it works.

I think that's super important because I think there is that very much the fear of missing out on this technology. And if you implement it incorrectly, you really are missing out. We have all the control functions reviewing what's being done and making sure that we're going the right way because that's the risk. And we can't afford that.

So we'll definitely go and go very slow and making sure that what we are doing is right. All right. Well, I thank you so much for your time. I hope that you have a great rest of the conference.

Thank you so much. Thank you, Lisa. Nice meeting you. And thank you for joining us on The New Age of Risk Analytics.

Please subscribe on your favorite podcast platform to be notified of future episodes and visit sas.com slash risk for information on these and other risk topics.

AnointedEDU Jermaine Whiteside The AnointedEDU Podcast explores the intersection of research, leadership, governance, education, artificial intelligence, and institutional innovation. Hosted by Jermaine E. Whiteside, doctoral researcher and founder of AnointedEDU, each episode translates scholarly research into practical insights for educators, ministry leaders, nonprofit executives, public-sector professionals, and organizational decision-makers. Through conference presentations, research conversations, working papers, NotebookLM discussions, and executive education, the podcast examines governance science, adaptive leadership, Faithful Intelligence, AI ethics, implementation science, institutional trust, organizational resilience, and evidence-based leadership. Whether you are leading a church, nonprofit, educational institution, business, or government organization, The AnointedEDU Podcast provides research-informed ideas designed to help leaders build resilient, ethical, Mavens of Data Maven Analytics A weekly live show bringing you expert advice and practical tips from top data professionals. Hosted by Maven Analytics. Daily Ops Boost Joshua Johnston Daily Ops Boost is your ultimate resource for mastering the inner workings of running a successful agency. This podcast is designed to redefine the way you perceive 'Operations,' going beyond systems and processes, and highlighting its pervasive role in every aspect of your agency - from lead generation, sales, and fulfillment, to HR and finance. Each week, we'll delve deep into strategic planning, client service, quality control, technology management, risk mitigation, and even cultural development to ensure your agency runs efficiently and effectively. Risky Business Scott Grandys Join Scott Grandys in the Risky Business podcast as he explores the greatest challenges facing the insurance and transportation industries – and the resulting innovations in risk mitigation.

Frequently asked questions

How long is this episode of Risk Analytics?

This episode is 25 minutes long.

When was this Risk Analytics episode published?

This episode was published on December 22, 2019.

Is there a transcript available for this episode?

Yes, a full transcript is available for this episode. You can read the complete transcript on the episode page.

Can I download this Risk Analytics episode?

Yes. Use the download control on the episode player to save the publisher-provided media file.