Skip to content
All writing

Fraud scoring inside a core banking system, step by step

Fraud scoring pays off most when the score reaches the transaction before it is approved. Here is how to get it there. When we put real-time scoring and alerts inside a major financial institution's banking systems, fraudulent transactions fell by 25%.

Banking & payments2 min read

Machine-learning fraud scoring works when the score is computed in real time on each transaction and acted on inside the banking system before the money moves. A score that lands in a separate dashboard minutes later only tells you which fraud you missed. Getting it into the transaction path is mostly integration and data work, done in the order below.

For a major financial institution, we built the feature store and chose the models against a held-out validation set. Scoring and alerts now run in real time inside its banking systems, and fraudulent transactions fell by 25%. The steps below are how we approach this work.

25%

fewer fraudulent transactions after machine-learning scoring went inside a major financial institution's banking systems

1. Find the decision point

Trace a transaction through the core banking and payment systems and find the last point where it can still be held, stepped up for verification or declined. That is where the score has to arrive, within the time the transaction allows. Everything else in the design follows from that point and its time budget.

2. Build a feature store

A model needs more than the transaction itself: the customer's recent pattern, the device, the beneficiary's history. Computing these on demand from the core system is too slow. A feature store keeps them current as transactions flow, so the model reads them in milliseconds, and the same definitions serve both training and live scoring.

3. Choose the model on held-out data

Compare candidate models on a validation set held out from training, drawn from a later period than the training data so the test resembles the future. Judge them on the trade-off the bank cares about: fraud caught against good customers stopped. A model that catches more fraud but blocks too many genuine payments will soon be switched off by the business.

4. Integrate into the transaction path

Call the scoring service from the decision point found in step one, with a strict timeout and a defined fallback, usually the existing rules, if the score does not arrive. Write the score, model version and reason codes onto the transaction record so investigators and auditors can see why it was held.

5. Route alerts to the people who act

Send held transactions to the fraud operations queue staff already work, with the reasons attached. Their decisions, confirmed fraud or false alarm, flow back as labels for the next round of training.

6. Monitor and retrain

Fraud patterns change when fraudsters learn what is caught. Track the hold rate, confirmed fraud and false positives daily, by channel. Retrain on a schedule and when the numbers move, and run every new model against the current one on held-out data before it goes live.

Questions

How does real-time fraud scoring work in a bank?
Each transaction is scored by a model at the last point where it can be held or declined, using current features from a feature store, and the result is acted on before the money moves.
Why does fraud detection need a feature store?
The model needs recent customer, device and beneficiary patterns, and computing them from the core system at transaction time is too slow. A feature store keeps them current for both training and live scoring.
What happens if the fraud model is slow or unavailable?
The call has a strict timeout and falls back to a defined rule, usually the existing rules engine, so payments keep flowing.

Tell us which system the AI has to run inside

If that system is not ready for AI yet, tell us that too. An engineer will reply and say plainly whether we are the right firm for it.