Sarthak Pattanaik is the Chief Information Officer of the Clearance and Collateral Technology group and is responsible for managing and enhancing the technology that supports Government Securities Services (GSS)’s Fed-eligible securities clearance, U.S. Tri-party Repo and Global Collateral Management products and services. Sarthak led the development and client rollout of Broker-Dealer Clearance (BDC), the next generation securities clearance platform. Additionally, Mr. Pattanaik has two patents for of blockchain and machine learning technology aimed at resiliency that support key production services. Prior to joining BNY Mellon, Mr. Pattanaik spent 13 years in Nomura Securities based in New York. He was the global head of market risk and counterparty risk technology responsible for value-at-risk, stress testing, economic capital and regulatory capital models. Prior to that he was a quantitative analyst responsible for the pricing of structured credit and securitized products. Mr. Pattanaik is a CFA and FRM charter-holder. He holds a Bachelor of Technology degree from the Indian Institute of Technology Kharagpur, and an MBA from the Indian Institute of Management Lucknow.View the profile
About the talk
BNY Mellon’s Government Securities Services (GSS) business is the sole provider of treasury settlement services in the United States of America. Given its unique market position, GSS is exploring how to help clients improve their forecasting of $70+ billion in daily settlement fails leveraging Google Cloud.
Speakers: Sarthak Pattanaik, Victor O'Laughlen
Google Cloud Next ’20: OnAir → https://goo.gle/next2020
Subscribe to the GCP Channel → https://goo.gle/GCP
event: Google Cloud Next 2020; re_ty: Publish;
Hi everyone. Mitchell Lachlan and sarta pattanayak here. Representing Bank of new-york Mellon, clearance and collateral management business. We're excited to what we've learned from our journey and helping solve a key financial services challenge. That has the potential to bring big-time benefits to our clients and the US Treasury Market is a whole Realize by leveraging Google Cloud. Powerful data science platform. As the global digital business leader for clearance and collateral management leveraging data through analytics. And machine
learning is a key Focus area for a business in order to enhance our ability to improve the client experience, weather is predicting a failed transaction or offering a prescriptive analytics solution to resolve it. Our future is firmly planted in operating as a date of first company has a long history of serving Appliance as a CIO in collateral, my responsibility is to focus on growing this business. What is important for us to know how we got here? And how the business give you to where it is today? It would share the importance of this exciting, new service that your building.
actually stars with the average investor who, for example, saves into a 401k If the investor rebalance is her 401k account, shifting from equities, in favor of bonds, then her 401K ministrator may get those bonds by contacting their trading desk who in turn goes to the bond market. And says, I'd like to buy treasury bond ABC for 95% of the face value. Buyer and seller named EB, make their offers. In a trade is arranged the buyer and seller agreed to swap the bond for cash. Notify a third-party service
provider of the agreement because they want that service provider to man is the actual exchange of bond for cash at the end of the day. the phone dead man will agree to many such trades throughout the day and most cases that exchange is successfully completed but sometimes it's not Party may not be able to deliver the bond due to some unforeseen event occurring after the agreement. But the investor still needs their bond. The fun dead men, keep some extra on hand to make
the investor hole in case of a settlement fail. How many extra should the admin keep well. That depends on how often a settlement of failure occurs. If the admin knew that they could right size, their stockpile, and free of valuable cash. This cash could improve Market liquidity by buying and selling more bonds or fund the development of new services for the investor. So, predicting settlement fails is key, but that would require a total view of the market. The
fund admin doesn't have that, but the clearing and settlement service does bny Mellon. Is that clearing and settlement service? Because we're in the center of them, are we have a total Market View and Wiz that to you, plus Google Cloud. Powerful analytics. We can predict which trades will fail to settle. Allowing the a man to predict how many spare treasuries to keep on hand. Of course by Mel and didn't begin the Cloudera. We had a 236 year Head Start. In 1784, we were
founded by then u.s. Treasurer Alexander Hamilton. Some of you may have seen the musical in 1789, the first loan to the US government was provided in the form of a treasury security. Treasuries are the largest and most liquid fixed-income Market in the world with b. Y melon still deeply involved in this market clearing and settling over 9 trillion and set a Google Securities a day. Because of our important role in the financial markets. A core mission of our business is to invest in resiliency and
value-added services. That help create efficiency and stability and helping solve the FED eligible. Securities fails challenge isn't a prime example of that. So we kicked off of Investigation or what we call. Formed a hypothesis into how we might help predict the settlement fails. The fails vary from day to day, but generally they amount to 122 per cent of approximately 4.5 trillion, and security, deliveries, and seems like a small number and it is. But due to the sheer size of this
Market, it amounts to approximately 70 billion of settlement, fails a day. Transaction that do not successfully settle by 3 p.m. Eastern Standard Time Each Day are considered fails, so we aim to predict the sales at 1:15 p.m. each day. Providing approximately 2 hours. Notice starter can you tell us how we went about this? The four stages to a machine learning process stage, one is focus on data ingestion, very build a data pipeline to move data from an
unprintable vaginal system to have an unplanned data Lake all the way to Google Cloud Storage. In the second stage, we're focused on data preparation. The first part of it is ensuring good data, quality controls and the second part of it is featuring Jhene. In the future engineering step. We start off with convectional information, things like account, security information, and transactional information. And then reiterate across a multi-functional group, get more information
from the clients from the business, from products, from developers and engineers. At the end of multiple iterations. We landed a P-51 video. 44 of them wear something that we identified as part of the decorations and most of these variables, that economic in nature, things like demand-supply variables, liquidity, variables velocity videos, The Purge stage is around model building. And then be a building, the model we were targeting three objectives functions. 1 accuracy, second performance, and third.
Interpretability and evaluated a whole plethora of models from simple regression to three base models, As we started with the models, we did finally settle with the light gvm model and the reason for that is the balance between accuracy and interpretability. Appliance repair interested to use the model which provides them directionality as to why a transaction has a higher chance of failure than another. The fourth stage is building the production pipeline
once he built a model which which would train and best having an automation, to take the output of the model, the output coefficients into a production system and ensuring that, that is seamless. What's critical processes. also important was the frequency of a 3D printer model every 3 months and also going to the model validation to ensure that the model did what is supposed to do What did we learn from this entire process? First is data collection. We are trying to transfer mod's from a transaction based organization to deter driven organization, we think about
the 5s with the volume velocity variety veracity in value. Velocity was the most important concept for us because we wanted to build a solid data Control process and solid quality. Controls up front, the second thing we learned was that under transactional systems, and storing everything before a lot on stirring, every half hour, I'm going to Ben's and restored the output, which is used for client reporting. Our end of the regulatory reporting. We try to store information at every point in time during the algorithm, for every decision
point that would attempt to look at the amount of credit usage, inventory, position and cash information. Because we do not know which of this data is going to be useful for our next experiment. Hardly realize very quickly with the success of this project has to be a mix of expertise from the data Engineers, as well as the subject matter. Experts As we were estimating, a couple of Sprint's 95% + accuracy. And then we went back and looked at it, reactivated found that we have played it
in the past. That was an interesting thing that the new model is significantly. Good. The second thing in this entire process that we learned was the concept of agility. This is a multifunctional team. And the goal of the team was to have small experiments in a piecemeal fashion and have an hypothesis for each of them. And as we address as we move forward or we go back and we start with a new set of sets of hypothesis. So for that to happen, it is important. That infrastructure is not a button, like an infrastructure is an enabler for that. I'm proposing that we
learned is digital transformation. This is not about technology, it's about people process and Technology together. When did Bill machine learning projects deterministic exploratory, please do not have the requirements document and design and development and testing process like in waterfall. You will have to think about small experiments. You have to look at hypothesis and you have to hit it so fast, it reaction is Key to Our Success. Now, let's talk about results.
Our Target was to predict 40% of the fields with 80% position and happy to say that we are able to hit those objectives. Not let me talk to Victor to talk about what's next in the service. Thanks. Arthur. Next is a client trial after approximately a year of development and testing. We are now currently in the process of trialing, the service with our key clients. This includes interacting with the service in a production like environment and evaluating the the benefits of the service so that they can trust that the service is accurate and consistent.
As the client provides feedback to us, we can seek out opportunities to make client specific model, optimization Are there, specific operational nuances about that particular client required in order for us to perspective in order for us to Precision a recall. Play edition, we will continue iterating on improving the service through Berry's Market Cycles. For example, in our line of business month and quarter and do your ends are important times for our clients and the Dynamics of trading change. Therefore making it more difficult for us to
predict settlement bills, but by identifying these different Cycles adjusting for the different training periods and looking at different model variables, we can find Opportunities to improve the model during those time. You finally, as we predict outcomes, we can that identify opportunities to have prescriptive analytics, integrated into those predictions to help mitigate the females which will then allow for clients to get more value out of the service. Finally headed over to sarthak to talk about why Google Cloud.
Why Google Cloud. The first I want to talk about White Cloud. If you think about this project, the two Focus areas, where around scalability and Agility the ability for us to run complex machine learning models, requiring High infrastructure, requirements, and if we have to do it on Prime, it would have taken a significant time as well as it would not have been cost efficient. So Cloud was a natural solution for this project. Why? Google cloud is Google bigquery. Big worry is highly scalable,
an extremely performing queries on our own private databases 2 to 3 days of transaction data. On Google Cloud McCreary, they were able to run the same set of three years at the transaction level and the performance was very similar to two to three days. The second thing was about collaboration. This is a perfect opportunity, for the Google engineers and the BMV melon Engineers to work together. Both of us, focus our own data. How do we build strong data practices in terms of
data ingestion? Did a storage indexing as well as delivering data to our clients. It was fascinating to watch what the team's come together and saw the business problem. Thank you for watching this and hope this was helpful for your cells.
Buy this talk
Buy this video
With ConferenceCast.tv, you get access to our library of the world's best conference talks.