Prescriptive analytics implementation for collections strategy: A Sequential Decision Modeling framework

Introduction

A B2B enterprise is characterized primarily by its need to rely on its customers or other businesses to promptly pay for the goods delivered.  In accounting terminology, we use the word Accounts Receivable (AR).  The process is straightforward.  A customer orders goods from the enterprise.  The goods get delivered to the customer (which is another business) and an invoice is delivered to the customer with a due date on it.  Typically, there is a reasonable amount of time between the time the goods were delivered and the invoice due date, to ensure the customer is satisfied with the delivery of goods.

In the ideal situation, we expect customers to honor the invoices, but that may not be the case and that is why we are devising a strategy to implement in the enterprise to improve the “dunning process”, which basically ensures that all the AR on the books are realized without delay.  

Predictive analytics framework

Our proposed framework introduces a predictive analytics model which we call the “Propensity to Pay” (PTP) solution.   We use the output from this model as input into a prescriptive analytics framework that will make the decisions on behalf of the enterprise.  

The predictive model is a classical AI/ML model that uses past history of customer behavior with many other attributes to determine the tardiness of a customer to pay on time.  

Based on our experience with building the PTP solution with our clients, we have observed that the model is highly accurate in being able to predict the right “payment segment” for every customer.   

The model allows us to enact a “proactive” strategy in our collections process.  Based on the predicted segments we adopt an engagement strategy that decides what messaging/channel is appropriate for communication with the segments.  This proactive strategy improves the likelihood of risky customers being engaged early in the collections process and to be able to reduce the Days Sales Outstanding (DSO).

Prescriptive analytics framework

Now that we have established our predictive analytics framework that characterizes the collection process, we now introduce the prescriptive analytics framework that will help us make optimal decisions pertaining to how to engage the customers identified in the segments.

The prescriptive analytics framework is an optimization framework that provides a solution strategy or policy that we can implement to realize our goals.  In our business problem, we need to make decisions over time and hence we have a notion of a time horizon (say, 7 days).  Since we have to make decisions  over time in a sequential manner, our predictive analytics framework falls under the paradigm of what we call Sequential Decision Modeling (SDM).

Sequential decision modeling (SDM)

The SDM framework undertakes to answer three main questions before we build the mathematical model to solve the problem.  The three main questions are:

1. What are the metrics that the business needs to focus on?
2. What are the decisions the business needs to make to realize the metrics?
3. What are the potential uncertainties or risks that can affect the goal of the business to achieve the desired metrics?.

Sequential decision analytics (SDA) provides a principled way to make decisions over time under uncertainty, incorporating learning and adaptation.  This is a superior approach to static decision modeling where business decisions are made as a one-time optimization that assumes perfect information and accommodates no feedback or correction.  However, in the real world, information arrives sequentially, outcomes are uncertain and good decisions made today shape better decisions made tomorrow.

SDA models the world as a series of evolving decisions under uncertainty.  At each time, we observe, act, receive new information and update the state.  The state is the information we have at a particular point in time.   A decision is an action taken constrained by the state.  Exogenous information is new information that we cannot control.  We use a transition function to transform the current state to a new state based on the decision made and the exogenous information at that time.  The goal is to choose a policy that makes good decisions that maximizes (or minimizes) a certain objective.  Typically, there are four classes of policies in the state-of-the-art literature in SDM.

What are the metrics?

Some of the metrics that we need to focus on are Days Sales Outstanding (DSO). DSO refers to the average number of days it takes a company to collect payment after sales has been made.  Clearly, any successful collections strategy would like to minimize the DSO. Another attribute that is of significance is Invoice Aging.

Delinquency rate is another metric that is of importance.  Losing customers is not a good thing for any business, but the business needs to make every effort to engage with the customer to ensure the customer is able to pay the invoices and maintain good credit with the business to ensure future orders for goods are made.

The collections department at any business consists of a number of employees that are focused on ensuring that the invoices are paid in a timely manner.  They interact with the customers by emails, calls or other channels and this utilization should be measured to ensure that all the employees are equitably engaged.

We are not interested in analyzing the quality of the call center operations.  But, we need to measure the efficiency of an employee in the collections department in terms of resolving unpaid invoices.

What are the decisions?

Typical to an SDM, decisions need to be made in time and the prescriptive framework needs to make decisions on when to contact the customer and the frequency of communication.  Certain employees may have relationships with particular customers that entails that only certain employees engage with a customer.  There are several ways of engaging with the customer through channels such as SMS, email, courtesy calls and letter of demand (LOD).  These decisions will be made in the prescriptive analytics framework.   The other important decision is how to prioritize the customers and typically in this case, the prioritization is based on Past Due Days and the amount on the customer invoice.

What are the risks/uncertainties?

Every prescriptive analytics framework needs to anticipate and address the risks and uncertainties involved that impact optimal decision-making.  In this business use case, some of the invoices may be delayed because of disputes around delivered goods such as quality, quantity and other attributes.  There is always a risk that some customers may default on their invoices making them delinquent.

The SDM framework is solved over a time horizon, so it is imperative that the model considers forecasts of future payments and purchases.   Any forecast is probabilistic and has its inherent risk.  The predictive analytics framework which is underlined by the PTP solution is probabilistic and there is the risk that the model may not define the segments incorrectly. Sufficient validation processes should be in place to check the risk associated with the predictive model.

Conclusion

We have introduced one of the most important operational issues in any business and that is to ensure timely collection of the Accounts Receivable (AR).  We present an integrated solution strategy that combines a predictive analytics framework with a prescriptive analytics framework.   The predictive analytics piece is centered on the propensity model.  We bring in Sequential Decision Modeling (SDM) as our approach to make optimal decisions in the prescriptive analytics framework.  We follow the fundamental paradigm on an SDM framework and that is to identify the metrics, decisions and risks or uncertainties.