How did a BNPL lender to increase payments 43% using Prodigal ProEngage?

ProEngage analyzes 500+ interconnected consumer account and interaction variables to deliver digital outreach recommendations for every customer. Here's how we fine-tuned it for one lender:

‍

‍

The challenge

Hi, I'm Akshat Vaidya. I am the product manager for the data science and the machine learning team at Prodigal.

So this large BNPL lender - essentially the world is moving on to a digital first strategy.

And that seems to be a wise move since the 7-in-7 on calls and the kind of restrictions that you have. But also in general, people are more likely to respond to text and emails. 

And this particular BNPL lender focused a lot on emails being their primary method of communication, primary method of engagement.

Now, the problem with emails is while they are cheap to send, they are also very cheap to ignore. 

Because you get so many emails all in a single day, and to just say, "I'm not going to care about it, I'm not going to open it, I'm not going to click it," is very easy.

So this particular lender had a problem where while they knew that email is the way that they want to move forward, they had no idea on what is the best time to reach a consumer.

And when you say, "I want to improve my email, click rates," because click rates is the holy grail of how you measure email performance, there are a lot of factors that go into it.

You can say you need to have the right message, you need to have the right visual feel of the email, you need to optimize it for mobile, you need to do a lot of things.

But the key thing for an email to be successful, what we have found is the highest lift, the highest impact thing that you can do is just send the email at the right time. Because if you don't send it at the right time, it is very easy for the email to get lost in the sea of emails that any person receives. 

The process

And hence the problem here was while they want to improve their click rates, the  recommended way or the highest impact-driven way was: For a given consumer, can you tell me what is the right time to send an email?

So at Prodigal, we are an interactions-first company. We kind of get signals from interactions. And here, interactions mean everything that you have - from calls, emails, texts, even static data for a particular consumer.

Where do they live? What was their balance? What was the time when they were put into collections? Their payment history. All of these are data points that while they are very valuable individually, the magic sauce comes in when you are able to connect and drive signal by interplaying and interlaying all of these different data sources together.

So now you have a clickstream data that is from emails, you have their calling data, which is from when they picked up a call or not, and then you have their payment history.

The special sauce or the secret ingredient here is our ability, Prodigal's ability to combine all of these data sources and convert them into features which the model can use.

The method

And here, the kind of features or the quantity of features that we have are so many that a typical sophisticated collections agency or a BNPL lender would have 50 or 60 parameters on which they will make a decision on when to call, why to call, what to do next.

We have the ability, because we are able to combine intelligence across all of these different sources to create more than 150 features. And the 150 features, essentially, for this particular lender, we created 170 features and used a model to kind of help them drive the impact that they were looking for. And the number 170 is only going to grow in the coming days. 

In fact, even today, we are constantly researching and building out what are new signals, what are new insights that we can find from these data sources and combine them together to truly give an omnichannel view of the consumer's journey, and based on that, help predict and help guide business recommendations.

Okay, so the result essentially is very simple and - simple to consume and operationalize. What we gave them was for every account that they wanted to send an email, we gave them that - if you want to send email to consumer Bob, the best time to send an email is 2:00 to 03:00 p.m. For consumer Alice, it is 6:00 to 07:00 p.m.

So we essentially gave them a 1-hour time slot of when will be the best time to send an email such that they will have the highest probability of clicking that email.

The actions

Now because of this information, what our customer here was able to do was they had a very static strategy of sending emails to consumers. They would say, all accounts that have a balance of $100, send them email at 03:00 p.m in the day. All accounts that have a balance of $200, send them an email at 05:00 p.m every day.

Because of our specific recommendation, they were able to personalize when to send an email. So now they can say that, "Hey, this person seems to be a working mom.

She probably does not really look at her phone in the evening.” For all consumers that are similar to this, we might be better off sending an email at, let's say, earlier in the day when the kids are at preschool or later in the day when the kids are asleep. So maybe send an email at around 7:00 or 08:00 p.m in the day rather than sending it - just blasting email at 02:00 p.m every day. 

What the consumer was also able to do, and this was proven by some of the testing that we were doing with them, accounts that have a certain delinquency age, specifically, we've seen that accounts that are younger in the sense that they are fresher placed, they are less delinquent, they often perform better.

But accounts that are 100 days, 200 days past due, they have a very low chance of - you have a very low chance of collecting on them. Because of this personalized and recommended and predictive email strategy that they were using, that was one that was powered by our model, because now the customer knew when is the best time to send an email to a consumer, we effectively saw that even these harder-to-collect segments of population responded much better.

So in the A/B testing that we did, the test population of these harder-to-collect 50-day-plus accounts performed really better than the control population of the same kind. 

The results

And this essentially means that because of our ability to segment, personalize and to drive this insight to the customer, they can now have very granular and hyper-personalized email strategy or just engagement strategies across the board, where they don't need to rely on archaic methods of segmenting their customers by one or two axes.  They can use all of the 170 features that we built to identify the best time and get better results. 

And one of the key things that came out of this email model, the email time slot prediction model that we had built, was that when we were looking at a month's or two months' worth of performance for this particular customer across the A/B test that we ran, we saw that on the test set of accounts, which is what our model predictions were being used for, we saw a clear lift, a clear better performance in both the open rate and the email click rate, which are what email performance is ultimately measured by.

And we were successfully able to show that by using our models, not only are we able to create signals, but also drive real business impact.

‍