How to Build a Highly Effective Data Science Program

Christine Hurtubise, vice president of product and data science at FIS Global, has a track record of building successful data programs as an early hire at some of the hottest emerging fintech startups, including Orum and Stash, as well as at established players such as FIS and OnDeck. 

I recently sat down with Hurtubise as part of the Tribe AI + AI Exchange Applied AI series, to talk about how to build effective data science programs at all stages of machine learning adoption—from zero to one, scaling early successes, and ultimately building a defensible data moat. (Spoiler alert: Hurtubise also shares what she views as the most impactful area for ML in fintech.) 

Here are the key takeaways from our Tech Talk. 

Create a Data Science Mission Statement 

This will save you time and drive success in the long run. Many companies rush to build when they should be laying the groundwork to define what they want out of data science.

“You have to think in layers,” Hurtubise said. “The first layer is the market: What is my company doing better, different, or faster than the market? The second layer is: How will data science enable us to do that better, drive more revenue, and improve our product?”

This reflects the experience we’ve had at Tribe. After running dozens of data projects across companies, we’ve similarly found that a planning phase takes risk out of development efforts and ultimately saves time.

From there, it becomes much easier to define a mission statement that answers the question: Why should data science exist at this company? Once you have that, you can figure out how to allocate resources for maximum impact, especially at an early-stage company, when resources are more limited. 

Start with a Proof of Concept that Can Influence Key Business Metrics  

Hurtubise joined Stash, a Series C personal finance startup, as the second data hire. She not only had to demonstrate how data could move the needle for the business, but also had to get organizational buy-in for building a data science program from scratch. The looming question was where to start.

“I thought product-first,” she said. “We needed to answer the question: What are the biggest gaps in what the product is able to do for the business?” 

At Stash, App Store and Play Store ratings were a huge driver of growth efficiency. And returned payments were one of the biggest challenges. Those were reducing the app’s rating, which in turn drove up customer acquisition costs—a metric that venture capitalists watch carefully as an indicator of marketing efficiency. 

Working backwards from the problem, Hurtubise dug in and saw that she had the data to solve the problem. The result was a 90% reduction in returned payments, a hugely impactful first proof of concept that not only solved a business problem but made a clear case for investing in the company’s data science capabilities. 

“Modeling products is labor-intensive,” said Hurtubise. “You need to be able to anchor it to an impactful area.” Being able to tie back to your company’s key performance indicators for confirmation is “a way to embed an internal feedback loop of demonstrating impact.”

Don’t Be Afraid to Hire or Structure Your Data Teams Differently 

You just may be surprised by the results.

Of the people I’ve met in this industry, Hurtubise is one of the most thoughtful about data science recruiting—and she’s done a lot of it. At Stash, she grew the data science team to 15 people, and she has hired more than 30 data scientists over the course of her career. This is especially impressive given that she takes a nontraditional approach to building teams. 

For example, on Hurtubise’s teams at both Orum and Stash, each hire did both analytics and ML. When I first heard this, I was stunned. It’s hard enough to hire ML engineers, so requiring data analysts to be ML engineers seemed like a bad idea. 

But her reasons for this approach are fascinating.

Hurtubise aims to make the technical skill sets interchangeable between the the data analyst and ML engineering teams, preserving interoperability and reducing organizational silos. 

This drives better data science outcomes in two ways:

• The best analytic studies use modeling approaches. For example, Hurtubise’s team may run a regression to quantify the impact of factors in a business trend, but it’s not always possible to isolate each change through an A/B test. In these scenarios, her team can run models to predict the business outcome on historical factors to offer a point of comparison to what happened (a pseudo control). If you don’t hire for this ability in your analytical team, you won’t be able to service these questions. 
• It helps in hiring the best talent if you can ensure candidates of a strong career path. Right now, ML skills are more in demand and command a higher salary. However, at companies such as Facebook, data scientists are strong analytical partners. By helping entry- and mid-level data scientists gain exposure to both skill sets, you can offer a more competitive career path.

Of the team members Hurtubise has hired, 50% have been women. I wish this weren’t a huge accomplishment, but it is. Some of her success can be attributed to doing the work to build a diverse candidate pipeline by partnering with organizations such as NYC Women in Fintech, Women in ML and DS, and Women Who Code. But Hurtubise also credits her success to being strategic about hiring for management roles first. 

“It’s easier to build a diverse organization when your management reflects that,” she said. “And this in turn naturally lends itself to eliminating some of the bias from your hiring process.”

Building a Data-First Culture Requires Different Strategies for Startups vs. Established Companies

Having a data-first culture means valuing data in making decisions across the company and as a product asset.

At startups, a data culture is key to driving the early insights that help you find product market fit. The challenge is to embed this in a way that supports quick-hypothesis testing and product iteration. 

At Stash, Hurtubise embedded data scientists into every product group so they’d be close to the product manager’s decision process. The process starts with analysis to see where customers are dropping off on the platform and to establish a hypothesis about why. The next step is to work with the product manager to design an A/B test, and the final step is to evaluate the test. The cycle moves the product incrementally forward to gain more customers. 

In contrast, established companies may already have large data assets, but monetizing them for new revenue streams requires investment in new technology. It’s important to create quantitative proof of demand across business lines to justify investment in tech such as MLOps infrastructure. Will revenue streams be at risk without it? Will costs be saved? Will we gain new customers? 

Tying the data to business metrics and projecting usage allows you to build profitably, which is key for adoption and trust. 

All this comes down to storytelling. Larger companies have decision processes that align with accounting and what they must report publicly to shareholders (profitability, cost per employee, market share). Startups are looking to maximize a growth story to the VC community. 

Teams of all sizes need to understand how the company’s story is evaluated externally and align with optimizing those metrics. Building a data-first culture is what helps you get there. 

Startups Have Technical Advantages, but the Potential for Impact at Large Companies Is Unparalleled  

Startups get a lot of credit for going all in on experimentation and being the most exciting place to apply data science. But Hurtubise and I both get deeply excited about established companies, with their troves of historical data just waiting to be unlocked. 

“An organization like FIS is really exciting as a data scientist,” she said. “It’s this incredibly data-rich company that wasn’t unlocking the potential in the way a fintech startup might. And there’s the opportunity to marry entrepreneurial thinking with the breadth and access that a company like FIS has.”

On the flip side, established companies exert tremendous pressure to match the scale of existing products and recognize revenue growth very quickly. There’s a tendency to want to go for products that already have product market fit, which lends itself more to acquisition than building in-house. But not taking advantage of the insights provided by existing data leaves a lot of potential value on the table.

There’s always the question as to whether to build your own intellectual property or acquire it as the most effective path toward growth and innovation, Hurtubise said. “Often, you can find places to leverage existing data in distinct projects. This is a place where working with external partners [can be really valuable]".

Finance-Specific Use Cases: Fraud Detection and Risk Management

Given Hurtubise’s expertise in financial services, it’s perhaps not surprising that she had observations specific to this sector. “At some point, every fintech company becomes a risk management company,” she said. 

Specifically, she was referring to fraud detection and risk management as some of the most impactful places to apply the power of data science for efficient decision making. In payments and lending, it allows you to give much more nuanced answers about degrees of risk.

“Essentially, we can use the data we have available at that moment in time to go beyond just a yes or no when it comes to risk and give a much deeper assessment of a person’s risk profile,” she said. 

This also helps drive access and affordability by allowing companies to scale out access to products and services without having to raise fees. But, as with so many of the other aspects we discussed, it all comes back to proving impact. 

Whether a company is making its first data hire, building a proof of concept, or looking to leverage a wealth of existing data, tying your data science and ML efforts back to business impact is essential. 

Learn More

• Watch the full conversation between Christine Hurtubise and Jaclyn Rice Nelson here: Leveraging Data Science–From Fintech to TradFi 
• How Waymo Is Using ML to Build a Scalable, Autonomous ‘Driver’
• How Data Science Drives Private Equity
• Building Autonomous Vehicles From the Ground Up

Check out Applied AI, a series of conversations around how ML is accelerating change across industries. You’ll hear from technical experts at top companies on how they’re using data to drive impact, operationalize ML solutions, and accelerate adoption across fintech, healthcare, investing, media, and more.