Part 1 of this article, discussed why and how Predictive Warning Systems should and can be used to mitigate risk. Now we turn to look at how Predictive Warning System’s can be built and implemented to achieve lasting and reliable results.
How can a machine learning model discover and apply boundaries between good and bad? The key is an intelligent exploration of the past.
Building a Predictive Warning System
You have an amount of past situations which either resulted in a bad debt or not. About these past cases, snapshots are taken at selected time points; at the moment of each snapshot there had been some particular recent history which we can paraphrase with an appropriate set of key performance indicators, also called “features”. Some features are likely to be applicable in any scenario, like the slope of the upward or downward trends in the 30+ and 90+ unpaid invoices. The fact, if there were at all any invoices going into 30+ or 90+? How the customer historically reacted on calls to pay? Whether the average outstanding of the recent months was considerably larger than the average amount a year ago?
In case of a B2C scenario you can exploit features related to the mindset of your customers, like what is their default payment method and in contrast to that what mix of payment methods they are actually using. In case of a telco provider, where customers could use a prepaid balance for using some services, a relationship was found between reliability to pay on time and playing lottery on the phone. In a B2B scenario, the amount of credit or debit notes, and their ratio to the overall balance was one useful factor.
There may be lots of other features which put a light on your risk exposure, all from a slightly different angle. If you have a rich enough dataset then a properly selected algorithm will be able to find the subtle differences between feature combinations which are rather positive or rather negative. And this is what we want – the desired model itself.
At the moment when you want to evaluate your momentary outstandings, the work to be done is much more simple. You just need to get your system to calculate all the needed features from the latest data. The actual features are substituted into the model – and there you are: you get your current risk assessment.The model building and scoring mechanism.
The volume of the final savings will depend on your business circumstances. But to mention a bright example, a department belonging to a large international telecommunication company has been able to halve its 90+ days outstanding volume within a few months with the help of a monthly run bad debt PWS, and the improvement went further in the longer run.
Be prepared and the value will be maximal
Invest energy into the below issues and you will establish a long-lasting and reliable bad debt PWS for yourself.
- Rich, consistent and reliable data: the most basic must•Precise definitions: what is really „bad” for you? Any debit balance? Or a balance over 5k EUR?
- Continuity: you need to have a consistent history and a consistent future. If your business changes fundamentally, then there are no permanent hidden logics to be discovered.
- Participation in the project: Do not be afraid to expect a series of meetings with your consultant, a number of unforeseen questions and small decisions. These will create the perfect fit between the prediction model and your real business.
- Allowing for imprecision: there will be cases where a statistical model will make a mistake – just like humans do. We are aiming at being significantly better in overall than simple rules, besides being much faster than unaided human experts, but not being precise on a 100% level.
- Training the people: the preparedness and buy-in of the people who have to use the predictions every day will be vital.
- Adjusting your non-IT processes: somebody is needed to read and understand the results at the proper point in time and be able to steer the actions. That person will need the time allocated for reading the model results and the entitlement to act upon the findings.
- Doing machine learning the right way: luckily, this is something that you can totally leave to the machine learning professionals!
- Predicting which partner debts may go overage or which accounts pose a risk of reaching a significant debit balance is a difficult task for humans. The reason is the co-existence of many business factors leading to a vast number of possible scenarios.
- Machine learning is able to provide data based models, which separate the harmless and the toxic outstandings from each other with high efficiency. Analytics experts can build models which learn from the past events the critical feature coincidences in an objective manner.
- The model can be put into the heart of an automatic bad debt early warning system which periodically scans all the outstandings for dangerous patterns. If any of those signs are found then it raises a warning much before a seriously bad debt is built up. If you trained your system like that, then you can have a preliminary assessment of the possible actions as well. So you can stimulate your collection processes in order to make preventative actions to save your money.
- As a result you can strongly cut the usual level of your aged debts and the write-offs
This article was written by Tamas Molnar , an advanced analytics expert who has successfully delivered analytical solutions in a number of industries to include telcomms, retail and banking, as well as across several domains (enterprise and residential sales, operations, marketing, finance and HR).