A password will be e-mailed to you.

Part 1

Issuing an invoice is one of the best things in the world. You have done your job well and now you are on the edge of collecting the fruits of your long efforts. You only have to wait a little and all your pains will pay off – except if the payment was late…

This unpleasant outcome may happen in B2C and B2B scenarios as well. People may simply forget about your invoice (in a B2C scenario) or a process can stop incidentally (in case of a B2B scenario). These cases can be remedied easily, but there exist more intricate versions. Some customers may be financially weak or untrustworthy. Some others may try to optimise their cash flow at your expense and retain the transfer. You may have a standard collection process which is triggered by your outstandings and spans from contacting the customer on multiple channels through suspending your services to legal actions or selling the debt. However these actions may happen too late. Or, if done automatically by a standard scheme, the strength and the cost of the actions may be inconsistent with the severity of the particular cases.

There exist analytical solutions which help by alerting you only if some invoices are really at the risk of becoming a bad debt in the future. Your outstanding items may be within the payment deadline now, but a well tuned advanced analytics system gives you the chance to highlight the risky items. For example they can help in the following ways:

  • Point out which individual invoices may age above some threshold (like 90, 120 or 180 days)
  • Looking at the account level instead of items, which accounts have a high risk to accumulate a material debit balance in the near future
  • Which outstandings may require just a notification to make the customer to pay and which ones need a stronger action

Due to their forward looking nature these tools can be termed as Predictive Warning Systems (PWS). They can contribute to the reduction of the outstandings and write-offs, because

  • They issue early warnings, so you have a longer time window to trigger collection actions
  • The warnings can be raised by the first precursors of disadvantageous trends at the partners, thereby enabling preventative measures, when still there is a higher chance to collect the outstandings
  • Resources can be optimally allocated to the cases requiring more expert attention
  • You can get advice on which possible measure to apply in a particular case.

How can you use a predictive warning system to reduce your outstandings?

With the help of alerts from a PWS you can eliminate a large portion of the hazardous items. The key point is that the warnings are issued quite a time before the situation would become unmanageable. The likely way of usage is periodical. With payment deadlines like 8, 15 or 30 days, you will not have fundamental changes in your claim portfolio within hours. Rather, your experts will look at the debts periodically and you will want to have an estimation of the outstanding risks at those points in time. The periodicity may be somewhere between daily and once in each few months.

Anything the frequency is, you accumulate information about financial events along some period and then use all your knowledge in a concentrated way. At that moment you can input the partner history data into a predictive system, which will inspect both the actual status and the trends which have led to the current situation. Then, adding everything together, the system will calculate for you risk indicators either for your individual invoices or your partner accounts, depending on the purpose for which it had been built. Your experts can then start elaborating preventive actions and execute them meanwhile all the new business events are recorded again and waiting for the next turn of evaluation.

When the present risk position is evaluated, then in the background the system aggregates together all the negative and positive trends related to your partner and expresses the overall judgement about the risk level in one main indicator. The model makes it possible to order your outstanding items into a list with the items having the largest chance to default on the top and the less risky items appearing below. A fixed line TV and internet service provider did exactly this when periodically scored those customers who had bought some precious equipment with part-payment. A model considered not only the timeliness of the installments, but also the payments against the service invoices and extra spending on on-demand services. The risk managers could go through this list and deal with the most dangerous items first.

You may have further aspects for prioritisation, like materiality or actual ageing. Any number of points of views exist, which may be subjective or sometimes contradicting, you will have a very objective measure of the hazards to look at first. Secondarily, you can get hints on the most efficient action to do.

You may have two sorts of risk mitigating actions:

  • Direct actions targeted at the individual hazardous items / accounts
  • Adjusting your general risk management approach as response to the discovered trends.
  • The first one is more concrete. If you see that some items have a high chance to go into default then you can take countermeasures without hesitation, thereby increasing the probability of collecting your money. If a partner account seems to be problematic then you can try both collecting your older outstandings and slowing down the rise of new outstandings. A large service provider company did the second: a direct channel was established between the revenue assurance and procurement departments and where both sorts of connection existed with an enterprise partner, the outgoing payments were delayed to counterweight the risk of unpaid invoices – even across countries. A retail bank chose to sell defaulted loans earlier if the chance to collect was very low, reducing the volume of the bad debt management activities.

The second type of actions means contemplating a bit about the results of the trend insights. Perhaps one market segment is just turning into a beginning recession, making multiple segment members hesitant to pay.The answers can be beyond your data, but the PWS will point to the negatively affected areas. Your responsemay be maintaining more cautious relationships with the affected segments.

Why to bother about building a predictive system? – Your motivation

The essence of your collection procedures is the set of decision rules when to do something and what. One way to create the rules is to collate and consolidate expert opinions on the important elements of non-compliant customer behaviour. In practice you will face serious decision making problems if you attempt to do this for a mature business:

  • The range of the possible cases is so high that humans are often simply unable to enumerate all possible situations and less able to assess the hazard level in particular situations.
  • In extreme cases it is easy to decide that the situation is almost sure positive or almost sure negative – but you have tons of cases somewhere in the middle where judgment is not so clear. A partner with invoices all above 120 days delay is very likely a lost one. But what about a partner which has been a good payer but only now some invoices have gone above 90 days?
  • You can easily be uncertain how the attributes of your partners exactly interact with each other. You may have multiple product types, several payment condition options, different customer segments. What ageing structure is perilous for one combination may be harmless for other combinations. In one case we found relationships, besides others, between the usage of self-service online channels and the timing of the payments: unfortunately the active (more informed?) users were more likely to pay slightly after the deadlines. On the positive side, in case of a small delay they were most likely topay by themselves and no action was needed.

The machine learning toolset is well cut out for helping with the immense complexity of the real-life business. The major goal of machine learning is to find valid, useful and interpretable interrelations within multidimensional data. This means discovering how all the business dimensions interact with each other and how the thresholds between good and bad can be found if all the intricate interactions are at play together. A useful result of the statistical exploration can be a model. The models express valid and comprehensive knowledge learnt from your past data, which is so generally true that it can be used to assess formerly unseen situations. Finally, you should have an automatic system implemented in your IT environment which employs the decision rules incorporated in your model and gives you the pleasure of exploiting its predictive power in an easy way so that one does not have to bother about the meticulous statistical work which led to the creation of the model.

Part 2 of this article, coming next Friday, will look at how to build out a Predictive Warning System and how to establish it as a reliable model.

This article was published 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).