Configurable Fraud Rules Engine with Versioned Rulesets
A fraud rules engine holds the conditions that decide which transactions look suspicious, and applies them to every event a monitoring system receives. Akoben's rules engine is the part of the platform that analysts work in directly. It evaluates each incoming transaction event against the active ruleset, records which rules matched, and passes the result to the scoring core and on to transaction monitoring, where matches become alerts. This page describes what a rule can express and how rules are changed, validated and reverted.
How rules work in Akoben
A rule is a condition on a transaction event, written as configuration. When Akoben evaluates an event (see the path an event takes), it checks the event against every rule in the active ruleset, notes which ones matched and why, and uses those matches to calculate the risk score and to raise alerts. The alert carries the rule or rules that fired and the specific values they matched on, so an investigator sees why the event was flagged without reconstructing it.
Rules are changed by analysts, not by engineers. There is no release cycle between deciding a rule needs to change and the change being in effect, and no ticket to a development team. The mechanism that keeps that safe, the ruleset, is described in the sections that follow.
Thresholds
A threshold rule fires when a single attribute crosses a configured value: an amount above a limit, a transfer to a destination outside an approved list, a withdrawal larger than a set figure. Thresholds are the simplest rule, and they remain useful because many fraud patterns do involve an unusually large or unusually specific value.
Their weakness is that they judge one value in isolation. Raise a threshold and genuine fraud slips beneath it; lower it and the alert volume becomes unworkable. That is why a threshold is usually combined with other conditions rather than used alone. Akoben lets a threshold be one condition inside a rule that requires several to hold, so the rule fires on a pattern, such as a large amount sent to a beneficiary added moments earlier, rather than on a number crossing a line.
Velocity rules
Velocity rules count events inside a moving time window: transactions per hour, cumulative value per day, distinct beneficiaries per week, failed attempts per minute. As events age past the edge of the window they drop out of the count, so the rule always looks at the most recent slice of activity.
The window length is what the rule is tuned against. A short window catches activity where the pace itself is the signal, such as scripted attempts. A longer window catches gradual draining, where no single transaction is large enough to trip a threshold but the accumulation is. Velocity closes the gap a threshold rule leaves: it looks at the same transactions in aggregate, where many small payments in a short period, or funds moving to a growing number of new beneficiaries, become visible.
See the transaction velocity rule in detail.
Transaction attributes
Transaction-attribute conditions test properties of the event itself: the amount, the currency, the channel, the type of transaction, the destination, the time it occurred, and whatever other fields the sending system supplies. They answer questions such as whether a payment is international, whether a wallet transfer is a cash-out, or whether it occurred outside the hours the account normally transacts in.
The attributes available to a rule depend on what the source system sends. A richer event gives analysts more to write conditions against, which is why the fields supplied during integration matter. A rule cannot test a value that was never sent, so it is worth deciding at integration time which attributes the fraud team will want.
Account attributes
Account-attribute conditions test properties of the account rather than the transaction: how old the account is, what type it is, what its status is, what limits apply to it, and what has been recorded against it. A payment of a given size means something different on an account opened last week than on one held for ten years, and an account-attribute condition lets the rule express that.
Account attributes are also where investigation feeds back into monitoring. A tag applied to an account during an investigation carries forward, so later transactions on that account are enriched with it and can be scored differently. A rule can therefore treat a transaction on a flagged account more seriously than the identical transaction on a clean one.
Behavioural conditions
Behavioural conditions compare a transaction with the account's own history. Akoben builds a baseline for each account from what that account has done before, and a rule can test how far a transaction departs from it. The baseline is per account, not a limit shared by every account on the platform.
That matters because a shared limit has to sit low enough to catch fraud on quiet accounts and high enough to leave busy ones alone, which makes it wrong for most accounts most of the time. A behavioural condition avoids the trade-off. An account that regularly sends large transfers is not flagged for sending another; an account that has never sent more than a small amount is flagged when it suddenly does, even if the amount sits under any threshold set for the wider population. A baseline needs history, so behavioural rules are usually combined with rules that apply specifically to new accounts.
Combining conditions
Real fraud rarely reveals itself through one signal. A rule in Akoben can require several conditions to hold together: a threshold, a velocity window, an attribute and a behavioural departure, in any combination the analyst chooses. The rule fires only when the whole combination is true.
Combination is the main tool for managing false positives. A condition that would raise too many alerts on its own becomes precise when paired with one that narrows it. A large transaction is ordinary; a large transaction to a beneficiary added an hour ago, from an account whose baseline shows nothing like it, is a different case, even though each fact alone might not justify an alert.
Rulesets: change, apply and revert
When an analyst changes a rule, Akoben creates a new ruleset. The ruleset is applied as a whole, as a single unit, rather than rule by rule. Edits do not reach live evaluation piecemeal; the new version of the ruleset replaces the previous one in one step.
If the result is undesirable, the ruleset can be reverted to the previous version, again as a unit. This is a different design from editing individual live rules. With individual edits, a mistake in one rule is hard to separate from the other changes made that day, and backing out means remembering and undoing each one. With whole-ruleset versions, the question after a bad change is simple: go back to the version that worked. Versioning also gives the team a clear record of what was in force at any point, which helps when a result has to be explained later.
Validating against historical transactions
A ruleset can be validated against historical transactions in batch before it is applied. The analyst supplies a batch file of past transactions, the ruleset is evaluated against it, and the results show how many alerts it would have raised, which transactions it would have caught, and which it would have flagged wrongly.
The rule definitions used in that batch run are the same ones that run in the live path. A rule proven against history is promoted without being rewritten, so there is no translation step in which the tested behaviour can drift from the live behaviour. For an institution moving from a batch-based monitoring system, this is the direct answer to whether its existing scenarios survive the move: they can be expressed as rules, proven against history, and then promoted.
Alerting-only mode
Each tenant can run a ruleset in alerting-only mode while validating it, then move it to acting. In alerting-only mode, the ruleset evaluates events and produces alerts and scores, and the team studies what it does without those results being treated as outcomes to act on. When the ruleset has proven itself, the tenant moves it to acting.
Akoben is not in the payment path, so the mode governs what Akoben recommends and raises, not what it executes against a payment. The decision to apply a recommended outcome belongs to the receiving system. The mode is a deliberate operating stage: validate in batch against history, run alerting-only, apply as a unit, and revert as a unit if it misbehaves.
Rule effectiveness and false-positive reports
Two named reports show whether rules are doing their job. The rule effectiveness report shows how each rule contributes: how many alerts it raised and how many of them led to a confirmed outcome. The false-positive rate report shows how often a rule's alerts turned out to be legitimate activity.
These are the reports a fraud operations team uses to defend or retire a rule. A rule that produces alerts that are almost always closed as legitimate is a candidate for tightening or removal, and a rule that reliably precedes confirmed fraud is worth keeping even if it fires rarely. Analytics and reporting covers the wider reporting set.
A module and a service
The rules engine is available in two ways, and both are real. Customers get the engine and run it themselves, with their own analysts writing, validating and applying rulesets. Akoben also offers rule design and tuning as a service on top, for institutions that prefer to have experienced analysts build and maintain the rules for them.
The choice between the two is a genuine decision. Running the rules in-house gives a team control and speed, but needs people who know fraud patterns and the time to tune them. The service reduces that burden, and an institution can start with the service and move rule maintenance in-house as its own team gains experience.
Common questions
What is a fraud rule?
A fraud rule is a configured condition that, when a transaction meets it, marks that transaction as suspicious. A rule can test a single value such as an amount, count events inside a moving time window, check attributes of the transaction or the account, or compare the transaction with the account's own history. A rule that matches contributes to the risk score and can raise an alert for an investigator to review.
Rules are written as configuration, not code, so the people who understand fraud patterns can express them directly. A rule has a condition, and the condition is built from the same few ingredients: a value, a comparison, and often a time window or a reference to the account's baseline.
Rules rarely work alone. A transaction that matches one weak rule is a different case from one that matches several, which is why rule outcomes feed a combined risk score rather than each acting as an independent verdict.
How do you set a fraud rule threshold?
Set a fraud rule threshold by looking at the distribution of real transactions, choosing a value that separates the unusual tail from ordinary activity, and then validating it against historical transactions before it goes live. The right value balances missed fraud against alert volume. Run the rule in alert-only mode first, review how many alerts it creates and how many are genuine, then adjust.
A threshold chosen by intuition tends to be wrong in a predictable direction. Set too low, it floods the investigation queue with legitimate activity; set too high, genuine fraud passes beneath it. Testing the proposed value against history shows both effects before any customer is affected.
A fixed threshold also has limits. It is the same for every account, which is why behavioural conditions, comparing a transaction with the account's own history, are often used alongside it.
Discuss your transaction monitoring requirements
Tell us which transaction flows you need to monitor, which systems Akoben would integrate with, and how your fraud team works today.
Request a demo