Introduction

Good evening, everyone. I'm Chloe Drew. I'm here to share how you can say yes to AI even in high stakes decisions and highly regulated industries, such as lending, the one I'm working in.

I'm a three-time founder in the fintech space. I've been, well, my full-time job is head of venture studio at EnsembleX. I've been an AI advisor to banks and fintechs for across 10 different companies.

Before EnsembleX, I was an operator at Mission Lane. It's one of the pioneers to use AI to expand financial inclusion for underbanked consumers. And before Mission Link, I was working a low-tech job called investment banking.

Goals for Today

All right, goals for today, three simple goals. As a practitioner of AI, I really want to exchange field notes, not just sharing them one directionally. I really want to tell you what I've seen, what I've observed, but also hear from you guys what you have seen and do you see anything different.

And secondly, if you're still curious about AI but haven't used it in your day job, I want to inspire you so that you can also say yes to AI going forward. And lastly, I want to make some friends.

I'll touch on how we can build partnerships a little later.

AI in the Financial Industry

So since we're talking about the highly regulated financial service industry, any guesses on what percentage of banks or financial institutions are using AI today? Any guesses? All right, we got two extremes. Two percent. All right.

But still, I believe there is a way to say yes to use AI, even in high stakes situations. And I've been doing that for many, many years to have my partners at EnsembleX. And we have helped many banks and fintechs to take that leap.

Using Common Sense

So here today, I want to share some field notes with you, again, to inspire conversations and maybe inspire some new users here. So first field note, very simple, use common sense.

Firstly, if there's an apple, don't reach for the orange first. What I mean is, if you want to predict someone's favorite dessert, don't start by looking at their favorite jeans brand, right? Look at what they ordered previously for their dessert, look at their whole food receipts, etc.

This phenomenon exists today in lending, whereas some lenders, especially fintechs, they would say FICO is crap. I would just look past FICO and use only alternative data. That's just reaching for the orange while you have a really good apple in front of you.

So use common sense. So what's the alternative to FICO? So there's a lot of alternative data such as trade line data that's traditionally not included or factored into the FICO score or cash flow data or text data which I'll touch on later.

Modeling Techniques

Next is use modeling technique if common sense prevails. For example, if someone has defaulted on every single loan they have taken out in the past, this person is way more likely to default on the next loan versus someone who has never defaulted on any loan in the past. So the likelihood just goes up based on the past trend.

It's just inconceivable if this trend would reverse, meaning if someone has defaulted on a 100 loans before and magically somehow their likelihood of default will go down over time, that's unlikely. So if that's the common sense, just put that constraint in your model. And lastly, if you can't explain a feature in English, be very careful using it.

That's because in lending we have this fun job of explaining no's, every single no to the regulator that's required by ECOA. And once I saw this model output from a data scientist, one of his top five features to predict someone's likelihood of default goes like this. the second derivative of a major purchase before a major holiday. So try to explain to the consumer why they're denied a loan because of this reason or also try to explain to the regulator why. So good luck with that.

Regulations

Next, speaking of know the regulations. Again, there's an alphabet soup of regulations. They're there for a reason.

If you operate in this industry, know them well. And better yet, try to work with regulators.

There has been successful examples in my industry. For example, Upstart is now a public fintech, publicly traded. They use AI also in their lending decisions.

And in 2017, they got this no decision letter from the CFPB. What it means is it's basically a get out of jail free card so that Upstart could use really interesting alternative data in their loan decisioning without any red flags or fines or even worse from the CFPB.

So if you do communicate and work with the regulators, you could actually open up new possibilities and be truly industry leading.

Complexity in AI

Next, less is more. So the complexity syndrome really shows up everywhere you look in AI. And I think this is one reason why AI has a bad rep, especially among regulated industries such as lending.

So on the technical side, it goes kind of like this. You probably have heard from a data scientist, hey, my model is amazing because it has a thousand features. That may be true in some other industries, but 1,000 features in lending usually means it's overfitted.

So the implications are real. If it's overfitted, you probably need to update your model every three, four months. A, that's a big operational lift for the data scientist, and also remember ECOA.

So you have to update all of your deny reasons every quarter, and then all of your operations team have to swarm, and your technical team have to recode all the reject reasons. So it's not stable from a model performance perspective. It's not scalable from an operational perspective, and it just leaves so much room for compliance gaps.

1So in fact, we actually worked with an auto lender here in the States. We cut their features by 90%, and the model performance actually was equally robust with only 10% of features left.

The complexity syndrome also shows up among business users. So if you work in lending, you probably have worked on rule version 78 in your decisioning. It usually happens like this.

As a lender, when you start, you have some simple rules and say, okay, rules-based or just deny and say yes in these situations. and then you launch your product. And three months later, a fraud attack happens and then you add some more rules.

And then another three months passes, a first payment default happens and you overlay some more rules. So over time, the rules get so convoluted, the data inside is actually trapped and the true predictive power of the data is not showing through.

So what we also did for a different lender is we actually peeled back their rules. It may sound scary, but by really applying AI gracefully, we were able to improve the predictive power of their model, staying compliant and making their life way easier.

Organizing Your Data

Lastly, probably the most non-obvious, organize your data. So everyone here wants to activate AI, use AI, and you probably have bought a few AI tools. But if you're dealing with AI day-to-day, you'll know that the data is still messy.

Conclusion

So that comes to the end of my presentation. I'd love to make some friends. Thank you.

Yes, I believe everyone can say yes to AI even in a highly regulated industry like lending. So if you are still a little nervous and you're new to AI, I'd love to help you take that leap compliancely.

On the other hand, if you're already an innovation leader and you're swimming in AI and you're already thinking about alternative data, cash flow data, text-based data and how they can be applied in lending in high stakes decisions to even better advance your business agenda and serve more consumers and getting more people into their dream homes. I'd love to talk to you too.

Some of the projects I'm really curious about is using cash flow underwriting for the credit invisibles and also using text-based data to provide additional lift to the predictive power. So if you're working in these fields, come talk to me later.

Thank you.