When Morpheus met Watson: JPMC teams with IBM Data Science and AI Elite
When the humans are away, will the AI play? It’s anyone’s guess. JPMorgan Chase (JPMC), however, is more than a little hopeful they’ll put their man-made-brains to placing the best investment bets.
It’s true. People no longer make the risky calculations that big banks teeter on; we now use financial models to roll the dice for us. Financial models are mathematical representations of a financial asset or portfolio created by algorithms, coded theories and AI. Quantitative analysts (quants) rely on these models to predict performance and make market judgments based on historic and financial data.
When financial models take the lead on important trades and investments, a new type of risk to manage surfaces: model risk. What happens if the financial model is wrong or inaccurate? Whether it’s badly fed by low quality data, poorly calibrated, or just plain old misunderstood—for the biggest banks, a poorly performing model can cause damages that swell into the billions.
There’s an urgent need and opportunity for big banks to mitigate model risk to avoid financial disaster. But they can also earn the trust of customers and leverage this position for a competitive advantage.
JPMorgan sharpens risk management’s teeth with data science
JPMC, one of the world’s largest banks, is on a mission to find optimal solutions to mitigate model risk by empowering quants and financial models with the most cutting-edge technology.
They looked to advance their model risk mitigation platform, called Morpheus, with a regimen of machine learning.
Partnering with the IBM Data Science and AI Elite Team, the IBM strike force of data science practitioners that help client organizations win with data and AI, JPMC made key enhancements to Morpheus to help manage model risk.
We sat down with John Thomas, IBM distinguished engineer and director in the Data Science and AI Elite Team, to discuss this unique partnership and his experience working collaboratively with JPMC.
Let’s dive right in: can you give us the 30,000-foot view of the engagement with JPMorgan Chase?
John Thomas: JPMorgan Chase is obviously one of the top institutions in the financial sector. We worked with them in a couple of different areas.
The one we worked on very recently is with their Quant Model Risk Management Group. They were looking to apply machine learning to help improve how their model risk management is done. Scenarios included classifying model usage restrictions automatically and predicting exposure of pricing models.
What value proposition did IBM Watson have to offer the JPMC risk management platform, Morpheus?
John Thomas: If you look at the data science workflow, especially in an enterprise setting, it's not just about building models. Anyone can go write some Python code in a Jupyter notebook. You've got to look at the end-to-end data science workflow – ingesting data, preparing data, ensuring there is governance around that data, and more – before you get to building and training machine learning or deep learning models.
Then what happens after you build those models? How do you actually operationalize those models? How do you compare model performance? How do you deploy them? How do you take the models from development to test to production? How do you maintain these models over a period of time?
With the Quant Risk Model Management use case at JPMC, one of the scenarios was around time series-based predictions. For example, can I predict exposures in these quant models? We need to deal with seasonality and trends in this time series data.
We got into building deep learning neural nets. When it comes to neural networks, training times can be pretty long. We had to cut down the training time for these models dramatically.
Let’s peek under the hood: what specific IBM technologies made this possible?
John Thomas: These enterprise challenges are met by a combination of Watson Studio and Watson Machine Learning to cater to the entire pipeline of the data science and AI workflow. We chose to run this on IBM Power Systems + GPU hardware that accelerates model training and allows us to scale model usage.
Watson Studio provides a collaborative environment where a team of people can come together to use open sources tools as well as value-added tools that IBM brings to the table and go through each of the stages of the data science workflow with the eventual target of operationalizing these models.
What good is a model if it stays inside your workbench?
You need to be able to get models deployed into production, scale them, monitor them, retrain them as needed—and actually be able to impact business using them.
That's what the Watson Studio, Watson Machine Learning tools, and the accompanying PowerAI platform, help with.
A common problem with data science projects is difficulty bringing all stakeholders into the fold. Is this something Watson Studio can address?
John Thomas: You've probably heard the expression that data science is a team sport.
Data engineering has to meet machine learning. Machine learning has to meet data visualization and that has to support the business user's interaction with the environment. Data science itself has to meet DevOps.
All of these different skills have to come together—and the platform makes it possible for individuals with different roles to come together and work on a project.
What are you most proud of when you think back to your engagement with JPMC?
John Thomas: What I'm most proud of with these engagements is how our teams work together.
The IBM Data Science Elite team and the client's team work hand in hand, collaboratively. This is not IBM going and doing some work for them and coming back to them after a few months. This is us working side-by-side, sharing knowledge, sharing expertise, and doing this in an agile fashion.
There is true value. Because we are learning from them, and they are learning from us, and together we are able to create something meaningful.