At this year’s North American Financial Information Summit, held on May 17 in Manhattan, machine-learning engineers from JPMorgan, BNY Mellon, and Morgan Stanley discussed where they’ve seen machine-learning (ML) projects go wrong. While some of the lessons learned from failures may seem obvious, inexplicable, opaque ML implementations still plague the industry. Financial professionals want to use ML as a magic bullet, but without a thought-out process and a plan, the end results often disappoint.

BNY Mellon’s Pablo Curello, who leads the portfolio of the bank’s ML and artificial intelligence (AI) initiatives in its Digital Group, kicked things off by detailing three common points of failure for firms when they begin their endeavor into emerging technologies:

“I would start with the user demand for the application, if you’re building a solution. I think many companies believe they have a lot of data they’re not extracting value from. Probably every company believes that, so you see it happen often that some groups put a team together, and they start developing analytics. They create dashboards. They create applications, and then after they put in a lot of work and it looks great, they start looking for who’s going to consume it. I’ve seen it a few times—you end up creating lots of dashboards, [and] analytics look great, but it gets very little usage. I think the first thing is to make sure that you source the ideas for the AI/ML use-cases from the actual potential users, and you ensure they actually want this, and you develop it with them.

“Another [point of failure] is the business case itself for the solution. There are so many things you can do with machine learning today. We oftentimes get new ideas for things we could work on, and we have to be very diligent about prioritizing what is worth doing and what is not because there are many use-cases that—if you really look at everything, you have to go through from idea to having the solution in production—are a real investment for the firm. … We do a lot of filtering for these ideas, and I think you have to get good at it. Otherwise, you can end up with a bunch of use-cases that look good but are not really making an impact.

“The last [point] is about scalability. I think you often see teams that develop an initial application that works, is good, and is getting usage, but they don’t really have a plan for how they’re going to scale it. If the goal is for the entire business—or multiple businesses—to use it, you’re going to have to think about it at the beginning. Maybe you have to do it on the cloud. … Sometimes you have to build admin tools because it’s not the same to have an application with a couple of models that a data scientist can train, as it is to have hundreds that need to be monitored and have to be re-trained up. You have to think about those capabilities ahead of time to make sure that you can scale the application to its real potential.”

JP Morgan’s director of data science Lory Nunez, who leads global AI initiatives in wealth management and private banking, advocated for building interoperability and holistic development into machine-learning models:

“Machine learning is not just about machine learning, so think of the problem holistically. It starts from use-case development, then your data, then your modeling, then the actual process. Once you think of it holistically, you have your use-case. There’s demand. You have the data, and you’re confident about your data, then you have the experts in the room for the modeling. Then you have to think, ‘How does this integrate into your process, into your operations, into your day-to-day?’ Once you think about that, you’re able to get feedback. You’re able to get a sense of ‘Can these people trust this? Can these people adopt this?’

“I’m also going to add interoperability. Machine-learning systems have various components, and we work in big firms. … Think of [machine learning] as component parts, where things can call on one another, and you build on the successes of another machine-learning team or maybe a vendor. How do you expose your machine-learning service so it’s interoperable?”

Morgan Stanley’s Suryakant Brahmbhatt, global head of data analytics, machine learning and visualization at the bank, said that in today’s age of information and data, the first question for an innovator is how to speed things up. But that mentality can lead to overly rushed productions, cut corners, and solutions built without the understanding of the data running through them:

“Our management has expectations that you will do something yesterday. We have expectations that we’ll deliver something today. But infrastructure partners support a month later. In this whole thing, we try to rush, cut corners, and what do we do? We don’t understand data at the level, and we build a solution. I’m seeing a lot of a new trend [wherein] the data scientist is sitting on the business side, and the ML, ops, data engineers, and data experts sit on the technology side. Before the technology equipment supplies the requirement to understand the data, the scientists want to rush. And what they’ll do is they’ll take Excel from somewhere, take in a dump of some data, and then they’ll come up with some model. And by the time we try to productize it, the data sitting on [production] vs. the sandbox is not seen.

“Sometimes the data for a couple years is pretty good. One of the areas we do is, say, how much should Morgan Stanley charge for a mortgage, or how much interest should they pay? We look at 30 years’ worth of data. If you build or train a model on two years’ worth of data, you will miss a lot of things. I think getting the data right up front [is important]. Normally in our area, we spend 75% to 80% in data engineering, and we’ve seen the result comes out very robust, very scalable, and it comes out very true to the question we’re trying to solve. So I would say spend more time up front. Don’t try to deliver something yesterday because none of us is Superman.”

One pain point all three agreed on when it comes to machine learning is transparency—or rather, the lack thereof.

Nunez, of JP Morgan, talked about the importance of treating transparency as a guiding principle, not as an add-on feature, in both data collection and sourcing, and model-building.

“We all think of transparency like [it’s] just putting a shop library there that’s going to explain the model. When you think of transparency and you’re at the model-building stage, you’ve got something coming. You have to think of transparency from the use-case gathering, where your business people have to be transparent to you on what outcome they want. What strategy is this outcome driving? You, as a data scientist, could make sure that you look at that data through that lens: removing any bias, making sure that you’re correct in your evaluation and assessment of that data, and getting the appropriate SMEs to help you evaluate that.

“Now, transparency from a data perspective—you make sure you know where your data’s coming from. What’s the data lineage? How good is your data? What’s the veracity of your data?

And then also from a process perspective, how is this model being deployed? Who are your users? Transparency in the whole lifecycle, transparency from day one and not just as a feature is what I would always tell my quants and data scientists.”

And BNY Mellon’s Curello built on the theme of transparency.

“What we do is we maintain a portfolio of initiatives in which we can easily explain what we’re doing in each one. Who is using it for what? What is the expected impact of the initiative? And we use that to communicate it to anybody within the organization, so people understand there are problems that we’re solving, how we’re solving them, and what it really means for the business. You could use metrics that come from the models, but usually the business is not going to care that much about that. You need to be talking about business metrics.

We spend quite a bit of time on each initiative trying to figure out how we’re going to measure that. If it’s an initiative that could grow revenue, how could we estimate what the impact would be ahead of time? And then how are we going to measure it afterwards? Is it sufficient? In each initiative, it really is kind of different. You have to figure it out. Data scientists usually are not very good at it. You need some consulting skills or business people involved who can help you do that. And once you get to it, then you need to be able to trace it. Every company is different, but in some cases and in some companies, people from finance specifically will ask for traceability—[for example] on which P&L is this sitting? Things like that. So sometimes you have to be very diligent about communicating and tracking all of that.”