The potential for artificial intelligence and machine learning (ML) is enormous. These new technologies, that work by crunching thousands of data points using mathematics that figure out structure and connections among every piece, have the potential to forever change the worlds of healthcare, communication and finance. But AI and ML ideas aren’t any good to anyone if they don’t make it out of the lab.
At ZestFinance, we’ve been actively using machine learning for credit and lending since 2009. Over the years, we’ve witnessed hundreds of ML enthusiasts successfully persuade decision-makers to adopt machine learning. Here are the practices they have in common.
Define the right problem to solve
Done right, ML is pretty powerful for credit underwriters — the main group of people who today use our Zest Automated Machine Learning (ZAML) suite of tools. But not every financial service problem can be solved by ML (as much as it pains us to say it).
Look at it this way: What ML is great at is hard for people, and what people are great at is hard for ML. So, if you have a lot of data that can be mined for new connections and insights, it’s likely that you can solve many problems using ML. For example, if a lender wanted to increase the pool of loan applicants, they could easily do so using ML to better understand the applicants’ data. Decision making around things that require human insight and interactions, such as hiring, aren’t served as well by ML. ML may help you sort a pool of applicants but, eventually, chemistry is going to play a big role in the hiring decision.
Make the business case for using ML
Deciding to use ML isn’t just about getting the technology right, it’s also about the time and expense involved. Show how the project will help the bottom line. Start with simple math — will the direct dollar benefits of ML be worth the direct costs of implementing it? In some organizations — especially in financial services — leaders often view cost-benefit through a lens of reducing risk, rather than just boosting sales. For instance, one Zest client considers its personal loan business riskier than its much larger credit card division, because they worry about single, large loans going bad. ML has helped make that division less risky.
Next, add in the carry-on reasons to adopt ML. What ripple effects and benefits will the increased automation have on the business? Is there an advantage for improved speed to market with ML? Is there a halo effect to be had from being at the forefront of technology? Many of these considerations resonate, especially with business leaders.
Build a coalition
You’re excited about ML — but not everyone else is. Accept that you will get some pushback from colleagues about ML because it is still a relatively new technology. So get a coalition to support the ML endeavor — business leaders, data science/model developers, risk management professionals, and the legal and compliance teams. Meet every stakeholder early in the process so they feel included and their opinions are heard. You don’t need to be an ML expert to explain to them what ML does. Call on your ML provider, if you have one, to join you in approaching your peers in a thoughtful and transparent way if you need help.
Practice like you play
In sports, you’ll often hear the phrase ‘you practice like you play.’ That’s a coach’s way of telling players to bring a serious approach from the start. Do that with your ML project and you’ll save valuable time. The first ML model you build probably won’t be the one put into production, but it could be. If you’re going to start with a proof-of-concept phase — and we think you should — operate as if you’re going to put that same model into production. If the model shows value, you’ll have avoided wasting another three months rebuilding something from scratch to get it into production.
Keep your allies involved. During model construction, meet with data scientists at least once a week and the general technology staff every few weeks to discuss implementation. Meet with legal and compliance early to keep them involved and then again when there is a model to review. Save time by meeting with your model risk management team while building the model, not at the end.
It’s vitally important that you know how your firm defines ML success. Understand which metrics matter to your company. When it comes to credit underwriting, for example, some banks care intensely about charge-offs, while others focus on increasing approvals. Then, set a reasonable timeframe to measure success. It’s not constructive to look at charge-off levels after six months of customer activity if your bank actually records charge-offs only after two years of delinquencies, for example.
It’s also important that you can explain how your results came about in order to bolster those success metrics. Too many ML packages are black boxes: data goes in and results come out but you can’t explain how the program got those results. Make sure your ML comes with true transparency, especially if you’re in the financial services world where you’ll need it to meet regulatory requirements anyway.
ML projects are as varied as the companies implementing them, but these five constructive steps are time-tested ways to get ML off the drawing board and into use. Once your model is up and running, the results will speak for themselves.