By: Hakan Karagul, Sr. Manager, Analytics
With self-driving cars cruising around, robots doing backflips and helping each other open doors, computers learning how to play GO in a few days and then beating experts who spent their lives mastering the game, we are definitely witnessing an exciting era in human history. Like Enova’s CTO John Higginson said in his blog post, as an analytics and technology company we want to use and even seek to extend these technologies to invent the future, but for credit. That’s exactly why our executive team picked ‘advancing our machine learning capabilities’ as one of our strategic initiatives this year.
Usually when people see these amazing advancements in technology the first thing they think about is how machines are taking over our jobs. In the lending industry, however, the takeover has already happened. Back in the day, people had to go to the bank and had to sit down with loan officers to get a loan (I know, hard to believe!). Loan officers used to make judgement-based decisions, which were not necessarily more accurate or more objective than the traditional statistical models (such as logistic regression and rule-based models) that almost all lenders are using today. Research has also demonstrated that people who make repeated decisions are not consistent with their decision-making processes and the outcome of their decisions can thus vary even at different times of the day. It has been shown that judges get harsher as they get closer to lunch time, for instance. So, replacing the loan officers with statistical models was a huge step forward: it made lenders much more scalable, accurate, objective and adaptable.
The companies that moved away from judgement-based decisions towards quantitative decision-making were the first to invent the future. The second breakthrough? Moving out from brick-and-mortar shops into 100-percent-online services. Once again, this was a huge step forward as it made lenders much more scalable and adaptable. Doing precisely that 14 years ago, Enova was one of the first lenders to offer online credit services by leveraging its advanced analytics and software engineering capabilities. In a sense, this was our way of inventing the future, for credit.
We believe that the next breakthrough in the lending industry will be driven by advancements in machine learning technology, and we want to be at its forefront. With more than 60 analytics professionals from diverse backgrounds, we certainly believe that we have the right skills and tools to make this happen.
Our first goal in this effort is to enhance our traditional models by using advanced machine learning (ML) algorithms like boosted trees, random forests and support vector machines with custom cost functions and constraints. Next, we want these models to improve their performance over time in an automated fashion. Last year, we proved the concept and demonstrated the value of these two objectives by creating prototype models using advanced ML algorithms as well as a prototype automated learning framework (aka ALF). Now, our next challenge is to take this same idea and apply it to all of our businesses in a scalable way.
You might ask what the big deal is about using these advanced algorithms and automating model retraining. Surely, this is all relatively straightforward to do on a local machine to solve a toy problem. Implementing ML solutions for real life problems have additional challenges, however. Here are a few of them:
Our analytics team is language-agnostic: we work with multiple languages and do not enforce a specific tool. Our machine learning platform should therefore be able to support multiple languages (R, Python, SAS, Mathematica) in a performant way. Designing and maintaining our model-running engine (aka Colossus) becomes a challenging task for this reason, especially considering the fact that we conduct nearly 1 million model runs every day and we don’t have the luxury to take it offline.
We make real time credit decisions within seconds. Fetching credit reports and parsing them, engineering new features, calculating scores, and optimizing the final decisions must be completed in this short time frame.
Just because we can pull training data from a reporting database does not mean that we can use all of those attributes in production. This is especially true in a microservice architecture, where each service might have its own database. Gathering data for model building purposes is usually an afterthought. We are changing that, however, and designing our decisioning platform with data availability in mind.
Dependencies Between Models
In most cases, our models don’t operate in isolation. They are usually part of a decision flow, such that the output of a model can be the input of the next one. In effect, optimizing and implementing end-to-end decision flows can be more challenging than optimizing and implementing a single model. Our new decisioning platform is also designed for decision flow optimization to meet precisely this challenge.
Legal & Compliance
We should be able to explain all the credit decisions that we make, and generate appropriate reason codes accurately for all the applications. This is a major prerequisite for our ML initiative. We made great progress last year by building custom tools to address these needs.
The Cost of Making a Mistake (e.g. overfitting, a bug in the model etc.)
A mistake can cost us millions of dollars. (No pressure.) We must therefore design robust model building and model validation processes as well as develop rigorous pre-release testing and post-release monitoring tools to make sure that nothing slips through the cracks.
As Enova’s data science team (internally known as RAP: Research Architecture and Platforms) we are excited to be one of the core drivers of our machine learning initiative–and, we are excited to address these challenges. We are confident that we can have a significant impact on all internal Enova brands as well as on external Enova Decisions clients. We have a lot to do ahead of us, however, so we need more hands on deck to accomplish our goals. If you are an analytics professional focusing on machine learning, or a software engineer who is excited to enhance our decisioning platform, check out our open positions here and take the first step toward inventing the future for credit with us.