Predicting accounts receivable payment behaviour
Case 12/16 of applied Artificial Intelligence
A client had already spent three years grappling with a problem. The client was a foundation that was to take over tasks from six municipalities in order to save costs. However, talks with the municipalities were not progressing very smoothly with regard to one of those tasks, managing bad debts. The foundation was meant to take over the debts, but how much were they worth? It was unlikely that all the debts would be repaid, so the amount had to be lower than the total debt, but the parties could not work out between them exactly how to determine that value.
“We were asked to create a dashboard showing the payment behaviour of all debtors up to now,” explains Wouter Pepping, Senior Manager of Technology & Data Risk at Deloitte. The request was to identify for each municipality and debt type the percentage of debtors that had paid off their debt in the past, and to deduce the value from that. “We proposed to take it a step further,” says Pepping. “By using advanced analytics, you can, in fact, make much more accurate predictions, down to the level of individual debtors.”
AI email alert
Get 16 AI cases in your email
A step further
The team tried out a number of models, and decided on a Random Forest model. “This is a machine learning algorithm, which uses data to train decision trees, and then creates a ‘forest’ of decision trees with random variables,” explains Pepping. “By navigating very large numbers of decision trees and allowing them to decide on the outcome, you get a close approximation of the risk of each debtor.”
This method is very effective when using inconsistent data sets, continues Pepping. “Two debtors may display the same payment behaviour over an extended period of time, but if one suddenly receives an inheritance or is promoted, that can change overnight. By creating a large number of random trees, the system is able to handle such differences more effectively.”
The result was a dynamic dashboard that can be sorted by different criteria, such as municipality, debt type, debt amount or period. You can zoom in closer and closer on the full list down to the level of an individual debtor’s transaction history. The system provides a prediction of each debtor’s future payment behaviour. “We validated the system and ended up with an average error margin of four per cent,” says Pepping.
An end to the ‘tug of war’
The solution provided the client with a breakthrough. The foundation’s interest was to arrive at the highest possible value of the debts, while the municipalities’ interest was to have them as low as possible. “Our method made it possible to determine the value with a very high level of reliability,” says Pepping. “It was considered a fair outcome by all the parties concerned.”
The model is broadly comparable with the risk analyses that banks perform on clients with debts, “but organisations outside the financial sector still rarely use advanced risk models, although they offer significant opportunities,” adds Pepping. “You can gain far more insight into data sets, which enables you to set your strategy much more accurately. And, as in this case, it could resolve a dispute after many years spent playing tug of war.”