Predicting Late Paying Customers - case study


Business Objective – Predicting late paying customers.

The Impact We Made - Achieved 87% accuracy in predicting customers that are high risk of paying late and improving cash velocity and days sales outstanding or DSO.

Summary - Early detection of late payment/paying customers

The goal was to leverage machine learning to create a predictive model that would help detect which customers are at a higher risk of paying late. By doing this, our client could better prioritize the collection agent’s time and put greater focus on customers at higher risk of paying late. The objective was to improve both the scalability of collections operations while at the same time increase cash velocity by reducing the average DSO.

About the Client – A leading software and cloud services company.

As their customers began migrating from on premise to Cloud computing the client needed to improve their collections and recovery process to create efficiencies at the scale necessary to support increased growth in their small and medium sized customer market. Immediate delivery of software and Cloud services also warranted faster credit analysis and improved collection decision making.

The Challenge – Mitigate growing collection costs while improving cash velocity.

The operational costs of collecting particularly at the volumes supported by the client were in a steady state of growth. Collections costs were largely linear to collection volumes. As receivables grew, the cost of collecting also continued to grow. The challenge was finding a way to mitigate cost increases by creating more efficiency in the collections process. The amount of outstanding receivables combined with the vast amounts of data available in the collections space, made this a great area of focus and opportunity for applying machine learning.

The Approach – Early detection of late paying customers.

Using a regression algorithm, we built a predictive machine learning model focused on the critical variables that proved to be most influential; in the case of our client, it was a customer’s geography/location, line of business and segment, types of products/services purchased, type and frequency of contact with the customer, age of customer, amount of invoice, variety/combination and quantity of products on the invoice.

One of the outcomes of creating a model like this is the benefits you get from data exploration and data insights. For example, in the process of building out and training this model we identified the key characteristics of late paying customers. Customers with long credit term extensions, the age of the customer (within 1 year), customers with specific products on their invoices, specific customer segments with payments due in specific months of year showed increased tendencies to pay late. Conversely, we also observed better collection results across customers with a high transaction concentration or customers with more frequent contacts before the due date where the contact was done either face to face or phone to phone. Out of these data insights and learnings we worked together with our client to create a set of recommendations and specific action plans to pinpoint improvements across invoicing processes, credit analysis, sales incentives and collections operations.

To achieve the benefits and business outcomes of machine learning, you need a good plan for how you are going to operationalize the model to truly influence and drive data driven decision across the enterprise. This takes an understanding of the enterprise system landscape and processes and is usually best done relying on a combination of deep system functional expertise combined with business subject matter expertise. We worked closely with our client’s operations team of credit managers and collections agents to create an operations plan for the machine learning model. Out of this plan, we implemented a very specific strategy where we embedded the machine learning model into the client’s existing line of business application (collections reporting tool), workflow/processes, reports and dashboards. The result – the client is now actively using the model day to day to help collections agents prioritize customers more likely to pay late and ultimately optimize their time and collections efforts putting improved focus on higher risk customers first.

The Outcome – Increased cash velocity and collections worth $20M annually.

The client was able to better understand the factors that influenced late paying customers the most and realign how collection agents prioritize their collection efforts and time. By improving focus and predicting late paying customers, the client accelerated collections across customers and created greater scale and efficiency with existing collections staff. Given the current average accounts receivable and customers volumes, the model is producing a conservative annual increase in collections cash velocity of $20M and growing. The business benefits are expected to continue to increase over time as accuracy improves with further data volumes and additional model training. While not included in the $20M annual benefit, our client is also expecting to achieve operational cost benefits (or cost avoidance) made possible with this predictive model.