New loss alert model offers more accurate risk differentiation for banks

VALID Systems introduces Real-time Loss Alerts, a new model designed to distinguish recoverable declines from true charge-offs, enabling banks to prioritise real financial losses and improve decision accuracy through expert-driven features and cautious bias mitigation.

In deposit-risk scoring, a decline is not always the same as a loss. That is the central point of a new VALID Systems paper by Michael Serrette, who argues that many bank decisioning systems stop at the moment an item is flagged without distinguishing between a recoverable problem and a true charge-off. The distinction matters because a returned item may be inconvenient, while a charge-off is money the institution may never see again. The model VALID has built, Real-time Loss Alerts, is designed to separate those two outcomes so banks can focus their attention on the accounts most likely to generate real financial damage.

Serrette says the model was built to sit alongside existing decisioning systems rather than replace them. Its job is to identify which declines are likely to become unrecoverable losses, enabling banks to hold funds longer, send cases for review, contact customers or continue monitoring. That focus on loss rather than simple return is important because, as Sardine notes in its explanation of charge-offs, a charge-off is the point at which a debt is treated as uncollectible and the loss is realised on the books. Sardine’s companion definition of loss rate also underlines why that distinction matters: loss rate measures realised losses after recoveries, not merely the volume of alerts or declines.

The VALID paper also stresses that the engineering behind the model is only part of the story. Serrette argues that the most valuable work lies in the features, the professional experience used to design them and the discipline around excluding inputs that cannot be justified. He says the model is built in the same environment as the data it uses, which reduces the distance between an idea and a test. That setup, he says, makes experimentation cheaper and allows the team to reuse feature interactions across clients rather than rebuilding them by hand each time.

Just as notable is the paper’s insistence on subject-matter expertise. Serrette says former tellers, credit union staff and mortgage specialists bring pattern recognition that cannot be derived from tables alone. He argues that many useful signals begin as an operational memory of how fraud or loss actually unfolds, then become measurable features only after that experience is translated into data terms. In practice, that means a model is not just a statistical machine; it is also a codified memory of how risk behaves in the real world.

The paper is also wary of bias, even when protected characteristics are not supplied to the model. VALID says it does not receive race, gender or age, but that does not eliminate the risk of proxy discrimination if geography, transaction behaviour or balance patterns reproduce the effect indirectly. The company says every feature has to map to a defensible loss mechanism, and anything that only works because it happens to correlate with a protected group should be left out. The article also says the team audits the finished model for concentration effects that could reveal hidden unfairness.

One of the more practical safeguards is manual whitelisting. Serrette says populations can change for legitimate reasons, such as seasonal labour flows, disaster recovery activity or government programme changes, and those changes may look suspicious in the data. Rather than letting the model penalise a whole group, VALID says it verifies the explanation and exempts the pattern. That approach is slower than automatic blocking, but the paper argues that restraint is preferable to building a model that accidentally treats ordinary behaviour as risky.

The other major theme is accessibility. VALID says the system was originally built for larger institutions that can provide a deep feature set, but it is now working on tiered versions for banks and credit unions with less data. Serrette is blunt that a model with only about 20 features will perform worse than the full version, but he says the challenge is to find the smallest useful configuration and identify which additional data points produce the greatest improvement. In that sense, the paper is less a product announcement than a case for a more cautious, more explainable kind of loss modelling.