Anticipating, 90 days in advance, which customers will pay without delay is a capability you can build into your operations today. It is neither a trend projection calculated in an ERP nor a gut feeling from the sales team: it is a figure, calculated invoice by invoice and rolled up to customer level, that is updated every morning.
Predictive credit risk indicators are metrics calculated with Machine Learning on each customer’s actual payment behaviour, which estimate what is going to happen (probability of default, payment quality, credit capacity) rather than describing what has already happened.
A well-calibrated predictive engine tells you which customers to keep a close eye on; it also tells you who you can sell more to, with how much credit and with what level of confidence. That is the shift in approach when the finance function stops being the brake and becomes the enabler of growth, using its own data.
Descriptive and predictive indicators: the difference that matters
| Descriptive indicator | Predictive indicator | |
| What it answers | What happened | What is going to happen |
| Horizon | Backward-looking: period close, month, quarter | Forward-looking, with an explicit time frame |
| Calculation basis | Aggregated accounting balances | Each customer’s payment behaviour |
| Useful frequency | Monthly | Daily |
| What it enables | Reporting and explaining | Deciding and anticipating |
| Examples | DSO, ageing, balance ageing analysis | Probability of default, behaviour index, calculated credit limit |
Both types are necessary and they do not compete. The descriptive indicator confirms; the predictive indicator anticipates. What you do need is for both to be reading the same portfolio: if you wish to fine-tune the descriptive side before taking the leap, start with how to read your DSO and release the cash trapped in your portfolio.
The three indicators of the Atlax 360 predictive engine
The Atlax 360 predictive engine trains its models on your own portfolio. It does not apply a generic market score: it learns how your customers pay you, which is different from how they pay your competitors or their bank. The result is three indicators that are recalculated daily.
- ICP · Payment Behaviour Index. Measures how, and to what extent, a customer is meeting its commitments in full and on time. It distinguishes the customer who always pays fifteen days late and with whom the business relationship works perfectly well that way, from the one whose behaviour started to change this quarter. It is in turn divided into three: how the customer pays your company, how it pays companies in your sector and how it pays all of its suppliers. All of this from day one, positive and negative and without discretion.
- RIIM · Immediate Default Risk. Estimates the probability that a customer will default on its invoices within a 90-day horizon. This is the indicator that provides the advance warning: which customers should be contacted before the due date and which do not need to be approached at all.
- LCA · Atlax Credit Limit. Calculates the maximum recommended amount for each customer at any given time, together with its associated level of risk. When that limit is written into your ERP or CRM and updated in real time, it stops being a monthly report and becomes a commercial lever: it releases orders without anyone having to interpret a spreadsheet.
All three then feed everything that happens downstream, from ongoing risk monitoring to collections prioritisation.
How to recognise a reliable predictive indicator
Not everything presented as predictive actually is. There are four signs you can check before entrusting a decision to a number:
- It is trained on your payment behaviour, not just on public data. A score built solely on information from company registers describes the general health of a business, not how it pays you. The most predictive data available on your portfolio is already in your collections history.
- It states an explicit horizon. ‘High risk’ is not an indicator; ‘probability of default within 90 days’ is. Without a time frame, there is no way of verifying whether it was right.
- It is recalculated daily. A customer’s profile can change in two weeks. A model that is updated once a month gives you a snapshot, not a signal.
- It is measurable. A model’s predictive power is expressed through the AUC ROC, a scale from 0.5 to 1 that measures how well it separates those who pay from those who do not:
| AUC ROC | How to read it |
| 0.50 | Equivalent to tossing a coin |
| 0.60 – 0.70 | Limited capability: useful as a reference, not for decision-making |
| 0.70 – 0.80 | Acceptable for supporting decisions |
| 0.80 – 0.90 | Good: sufficient for setting limits and priorities |
| Above 0.90 | Excellent, although it is advisable to check for data leakage in training |
Atlax 360’s risk models operate at an AUC ROC of around 88%. Always ask any provider who talks to you about prediction for this figure: if they have not measured it, they are not measuring anything.
Where they are applied: five decisions that change
An indicator is only as valuable as the decisions it enables, so here are the five where the return appears first:
- Onboarding new customers. Transaction evaluation no longer depends on a weekly committee and is instead resolved on the spot, with consistent and traceable criteria.
- Setting and reviewing limits. The most profitable case is not lowering limits but raising them in good time: customers who have already shown good behaviour and to whom you are selling less than you could.
- Portfolio segmentation and alerts. The team no longer reviews the entire portfolio in the same way and receives alerts only when a customer moves segment.
- Collections prioritisation. The order of contact is set by risk, not by the age of the invoice. This is the foundation on which any collections automation then operates.
- Provisioning and expected loss. With a probability of default for each customer, the IFRS 9 impairment model is based on proprietary, auditable data.
How to get started, in four steps
- Gather your collections history. Two or three years of invoices, with issue, due and payment dates, together with kept promises to pay. You need good raw material for the model.
- Define the horizon and the decision. Choose just one: limits, onboarding or collections prioritisation. Let us start with one and then move on.
- Train and validate with a real portfolio. Compare the prediction against what actually happened in a period the model has not seen. That is where the AUC ROC comes from.
- Integrate the indicator where decisions are made. A number that lives in a report changes nothing. Written into the ERP or CRM, it changes the next day’s operations.
Any questions? That is why Atlax 360 offers a simple implementation in 8 weeks, so that you do not have to do it yourself, combined with behavioural data from other suppliers and public information.
The predictive engine at the core
Everything an Order to Cash platform does well (automating, prioritising, scaling, provisioning) depends on the quality of the indicators underneath it. That is why, at Atlax 360, the predictive engine is not just another feature: it is the layer on which everything else rests, including the Agentic Collections module, which takes action based on what the engine calculates.
That foundation is built on volume: more than 800,000 tax IDs (RUTs) analysed in Latin America and more than 2 million debtors in Europe, across more than 2,000 projects. Companies working with Atlax 360 record an average DSO reduction of close to 10.5%.
You can see how this translated into three different operations in LATAM Airlines Group, Sika and Correos Express case study, or find out how Atlax 360’s AI and Data Cloud work.
Frequently asked questions
How do they differ from a traditional credit score?
A traditional score is built on public information and describes the general health of a company. A predictive indicator is also trained on your collections history, so it measures how that specific customer pays you. Two suppliers may have different risk profiles for the same company.
How much data do I need to get started?
Between two and three years of invoicing and collections history is usually enough to train a model with measurable predictive power. The quality of the records (actual payment dates, kept promises to pay, incidents) matters more than the absolute number of customers.
What is the AUC ROC and why should it matter to me?
It is the standard measure of a model's predictive power, on a scale from 0.5 to 1. Between 0.80 and 0.90, the model is good enough to set credit limits and collections priorities. It is the figure you should ask for from any provider who talks about prediction.
Do predictive indicators help to sell more, or only to avoid defaults?
They help with both, and the commercial return is usually greater. Identifying customers with good payment behaviour whose limit you can increase releases sales that are currently being held back out of caution, without increasing actual exposure.
How do they integrate with my ERP?
The indicators and limits are written directly into the ERP or CRM and updated in real time, so that orders are released or held within the system your team already uses, without switching screens or consulting a separate report.
Does it work for a group with several subsidiaries and countries?
Yes. It allows a consistent corporate credit policy to be applied while respecting differences by country, business line or currency, with comparable indicators across subsidiaries and consolidated real-time reporting.
