Reading payment flows without moving money: what anomaly detection says and does not say

The duty to watch

Paragraph 11.6 of PSD-12 requires entities in the National Payment System to establish capabilities to monitor and detect anomalous cyber activity continuously, including the monitoring of all payments for fraudulent or suspicious activity. The Bank of Namibia’s fraud report of March 2023 gives the reason. Between 2013 and 2022 it counts 66.2 thousand incidents worth N$158.9 million, and it sets a system-wide fraud threshold of 0.05 percent for continuous watching.

Watching at scale is a data problem before it is a model problem. A provider sees only its own customers. A switch or a central bank sees all of the traffic, but sees it as messages, with few of the facts about who the parties are.

Observer, not participant

Two kinds of system need separating. A participant starts, authorises, holds, reverses or settles payments. An observer receives information about outcomes under an agreement and analyses it. The observer is unable to start a payment, stop one, hold a balance or change what the rails recorded.

The difference is more than a label. It shapes risk, because an observer with no authority over money has a much smaller attack surface and a much narrower legal question. Whether an observer-only data relationship needs authorisation under the Payment System Management Act, 2023 is a question for the Bank, and the right course is to ask it directly instead of assuming a favourable answer.

Borrowing a method honestly

The method literature is useful. Ajit Desai, Anneke Kosse and Jacob Sharples published a machine learning framework for anomaly detection in payment systems, issued as BIS Working Paper 1188 and as a Bank of Canada staff paper in 2024. Its abstract describes real-time transaction monitoring in high-value payment systems, a central piece of a country’s financial infrastructure, which system operators and overseers use to detect anomalous transactions caused by a cyber attack or an operational outage.

The framework is layered. In the first layer, a supervised algorithm identifies and separates typical payments from unusual ones. In the second, only the unusual payments pass through an unsupervised algorithm for anomaly detection. The authors tested it with artificially manipulated transactions and payments data from the Canadian high-value system. They report a detection rate of 93 percent for the first layer, a significant improvement over commonly used econometric models, and the second layer marked the manipulated transactions as nearly twice as suspicious as the originals. These are the authors’ results on Canadian data. Namibia’s everyday rails are retail rails. Payments are small, parties are many and the patterns of a market stall differ from those of a clearing bank. The method’s logic travels. The figures do not, and any figure quoted for Namibia must be computed on Namibian data.

The ground truth problem

No public data set of confirmed fraud exists for Namibian payment rails. Without labelled cases, a team has two honest options. It deliberately alters real payments and tests whether its models catch the alteration, comparing each altered payment with the original it was made from. Or it says its output is a ranking for human review, not a verdict.

The first option measures the pipeline, not the world. A model finds an alteration someone planted, and finding a planted case says little about finding the next real scam. Every detection figure produced this way is a test of plumbing. Thresholds would have to be derived again on real data before anyone relies on them.

Explainability is part of the product

A flag with no reason is hard to act on. Each flagged payment should carry the rules it hit, the features which moved its score and the model version, so an analyst sees why it was flagged and rejects it with a reason. The draft Data Protection Bill gives individuals a right to know the reasoning behind processing whose results are applied to them (section 8(4)), and a rule-based or learned score applied to a person falls within this principle.

Measuring honestly when there are no labels

The paper’s own design is instructive. Because confirmed anomalies are scarce in a high-value system, the authors test their layers on artificially manipulated transactions, then compare each manipulated payment with the original it was made from. The comparison is paired, so a score for the altered payment is judged against the score of the real one. A model which fails to separate the two has learned nothing, and a model which succeeds has shown a specific skill.

A Namibian team without labelled fraud has the same option. It alters real payments in controlled ways, such as amounts, timing or counterparties, and check whether its models rank the altered payments above the originals. This tests the machinery, and it does so honestly, provided the report says what it tested. A result of the form: our models rank planted anomalies higher is useful for debugging. A result of the form: our models detect fraud is a claim about the world which the test did not examine.

What a reviewer should see

An analyst who receives a flag needs three things. The first is the reason: which rules fired, which features moved the score, and the version of the model. The second is context: the history of the parties and a comparison with similar payments. The third is a route: a way to confirm, dismiss with a reason, or escalate. Without the third, a flag is a notification and not a control, and a review queue which grows faster than it is cleared is a sign of a threshold set too low.

Dismissals matter as much as confirmations. Each reason recorded by an analyst is a label for the next version of the model, and it is the only ground truth a retail system in a small market is likely to gather. A programme which stores analyst decisions in a form a model learns from will, over time, replace its planted tests with real ones.

Why Namibian scams defeat a size filter

The Bank’s fraud report shows why the detail of retail fraud matters. Phishing accounts for 92.5 percent of electronic funds transfer incidents, and phone call scams dominate e-money fraud. In both, the customer sends an ordinary amount to a new payee under pressure. A monitor tuned to large or unusual amounts misses the case, because nothing about the amount is unusual. Useful signals are the novelty of the payee, the speed of the sequence, the time of day, and changes to the customer’s device or SIM shortly before the payment, since the report lists SIM card swaps among EFT and e-money fraud types. Each signal needs data a provider holds and a switch does not, which is why an observer and a provider work best together.

What this means for organisations in Namibia

A bank or provider building monitoring should write down its ground truth plan first: where labels come from, how late they arrive, and who reviews flags. A monitoring model with no review queue produces noise.

A supervisor buying analytics should require a clear statement of what the system is unable to do, a record of every figure’s data source, and evidence of how a flag is explained to a reviewer.