Improved Collections

Overview

Each borrower is unique and so are the reasons they may default on their loan. Circumstances like divorce, illness, job loss, failed business, and over extension of credit are some of the common, yet vastly different, reasons that borrowers default on their loans.

The approach to collections for a newly past due account with a good credit score and a long history of on-time payments should differ from the approach for a newer account with several missed payments and other high-risk factors.

What if you could automatically tailor your approach to collections based on the unique circumstances of each borrower?

This Chicago-based consumer finance company partnered with 2nd Watch to build a machine learning solution for improved collections to analyze unique borrower attributes, define segments, and ultimately enable a more targeted collection strategy based on the circumstances of the borrower.

The Challenge

Historically, our client had been reasonably successful in their collections strategy by segmenting and targeting delinquent borrowers solely by the number of days they were past due. When the client realized that changes to their approach could lead to nearly $1M annually in additional collections for each reduction in their net charge-off rate, they sought out a solution to make that happen.

Their goal was to develop and implement a data science strategy that could:

• Analyze different attributes and dimensions, including demographics, preferred contact methods, past payment history, geography, etc.
• Define borrower segments based on common attributes and circumstances.
• Use machine learning and predictive modeling to define and deliver a custom communications strategy for each segment.

The Solution

Working closely with the client, the 2nd Watch data science team developed an advanced analytics and machine learning solution to make it happen.

Step 1: Data Collection and Defining Characteristics

2nd Watch’s data science team started by collecting account information such as loan characteristics, borrower demographics, and historical transactions to generate a rich set of descriptors. Our team then pre processed the data by ensuring validity and consistency of data points and used the raw data to engineer highly informative measurements.

Step 2: Calculating the Risk Score

Using a supervised machine learning technique, this data was correlated with historical default information to generate a risk score for each borrower.

Step 3: Borrower Segmentation

Using open-source machine learning implementations and cluster analysis, the risk score was combined with additional indicators to identify segments in the borrower base.

Step 4: Defining the Profiles

Utilizing unsupervised learning, 2nd Watch extracted seven representative profiles (prototypes) of borrowers based on which collection strategies could be devised. Each of the borrowers was then assigned to the prototype it resembled the most.

The Outcome

The seven borrower prototypes empowered the client to customize collection strategies based on each prototype’s unique circumstances. The collections team was able to optimize communication schedules, tactics, and messaging to increase the chances that their collection efforts would be successful. They were also able to use this information to prioritize higher-risk prototypes and provide context for their clients. The net result of this machine learning solution for improved collections was a reduced net charge-off rate with a significant improvement in profitability.