Earn Machine Learning Buy-In From the Right People

For many lending executives, machine learning is top of mind these days. ML has the potential to build supercharged credit models that can result in higher lending volumes and lower default rates by using high-level math to make billions of calculations, finding unseen correlations among hundreds of disparate pieces of borrower information.

It’s no surprise, then, that a Fannie Mae survey of 184 lenders last year found almost 60% expect to be using ML by the end of next year. The survey found most credit unions are taking tentative steps toward ML, with their efforts about equally divided among three areas: Rolling out ML on a full or trial basis, actively planning ML usage or taking a wait-and-see approach over the next year. In my over a decade of experience deploying AI and ML in finance, the biggest delay credit unions will face implementing ML isn’t technological – the tech is simpler than people think. The real challenge is getting buy-in from the people in your institution who can make or break the project.

Since ML is just starting to be adopted in credit underwriting, a lack of familiarity triggers fears of the unknown. Some people worry ML will automate them out of a job, while others fret that the monetary cost of moving to ML is too high. These fears tend to be greatly overblown. To get an ML project approved, you likely need to convince six key people that the effort is worthwhile, that it will improve the bottom line, and most importantly, that while jobs may change they won’t necessarily be lost.

The Business Executive

Improve the bottom line while being forward-thinking? Every executive wants that. Luckily, the business case for ML is pretty easy to make. In many cases, lenders I’ve worked with have seen approvals jump 15% or charge-offs drop by 30% after switching to ML underwriting. No matter what the specific numbers are for you, ML can cut expenses by reducing credit union exposure to bad borrowers and setting more accurate prices.

The Credit Risk Manager

The credit risk manager is usually the one responsible for profit and loss in the underwriting business. They’re going to be nervous about the “black box” reputation that plagues some ML systems that can’t explain how they reached a specific credit decision. Risk officers have to know the decisions are accurate, consistent with policy and non-discriminatory. They need transparent ML that includes reports that audit every decision, so make sure your ML system provides that.

The Credit Modeler

Many modelers spend countless hours working on proof-of-concept models of all stripes that never came to fruition. So it’s understandable that they’d be hesitant to adopt something as seemingly complicated as ML. But the truth is that most successful ML programs rely on a trusted outside vendor that will work in partnership with in-house modelers. Another plus is that some vendors offer automated documentation that produces reports as the model is being programmed, saving the modeler many hours of tedious documentation work.

The Model Validator

With an average of about 30 variables in a traditionally built model, model validation is a difficult but manageable task. The fact that an ML model could contain thousands of variables will likely strike fear into your model validator’s heart. But ML doesn’t require manual validation. A well-built ML system will come with explainability tools that detail how the model makes its decisions. Those tools make the process of validating an ML model straightforward, closely resembling the governance models the team’s already been using.

The Regulatory Compliance Officer

This person’s worst nightmare is scrolling the morning’s news and seeing that their institution is being accused of discriminating against borrowers. These fears can be exacerbated by ML’s complexity. The compliance officer is worried about hidden bias violating the credit union’s Fair Credit Reporting and Equal Credit Opportunity obligations. Assure them that the ML model’s built-in explainability will show why decisions are being made in a straightforward manner, and that any bias will be surfaced and can be removed from the model.

The Head of Information Technology

Pity the IT chief, who is always being asked to do more with less. When you say “machine learning,” to them it may sound like “really big complicated project.” IT heads are leery of committing to any project that will take months or years and require lots of staff. But ML shouldn’t take that long. Explain your ML developer’s track record of getting developments done in a timely fashion (in some situations, just weeks). It will help your case if your ML partner can provide supplementary staff support during the model implementation process too.

The bottom line is that your credit union’s processes don’t need to change much with ML and people won’t see their jobs automated into something they didn’t sign up for. Approaching colleagues with confidence and understanding will go a long way toward winning allies. The challenge then is to make sure you choose the ML approach that will make your promises come true.

Jay Budzik is Chief Technology Officer for ZestFinance. He can be reached at partner@zestfinance.com.