Venor MarketsMarket PulseHome

Market Abuse Surveillance Tuning: How Brokers Calibrate Automated Alerts and Manage False Positives in Order and Trade Monitoring

Why Your Bank's Abuse Alerts Miss the Real Crooks

Brokers' automated surveillance systems flag thousands of harmless trades. Calibrating them well is harder — and more important — than it looks.

Published · 2026-11-19

Picture a smoke alarm. It goes off when you burn toast, when the shower steams up the hallway, and occasionally, mysteriously, at 3am. You learn to wave a tea towel at it and carry on. The one thing you must never do is assume it will always shout when there is an actual fire.

Brokers and investment banks run something similar for market abuse. Under the Market Abuse Regulation — the rulebook banning insider dealing and market manipulation — firms must have effective systems to spot suspicious trading and report it to the regulator. In the UK that means the Financial Conduct Authority, and the rules are known as UK MAR. (Slightly different versions apply in the EU and the US, so jurisdiction always matters.)

Those systems generate alerts: automated flags saying "look at this order" or "look at this trade." The trouble is that most alerts are not abuse at all. Some observers put false positives at 90–95% of the queue. That is the toast problem, at industrial scale.

What an alert actually is

An alert is a rule meeting data. A firm configures a scenario — say, "flag a client who buys a bond shortly before a sharp price rise" — and the system scans orders and trades against it. Each scenario rests on parameters: a price-movement threshold, a time window, a repetition count. Calibration is choosing those numbers.

Set a threshold too tight and analysts drown in noise. Set it too loose and genuine abuse slips silently underneath. The FCA's Market Watch 79 notes that failures can leave an entire slice of business unmonitored, leave a scenario partly working, or leave it generating nothing at all.

Here is a real example from that newsletter. A firm adopting a new third-party system needed two inputs for its insider dealing model: a significant price move and a news release. The news feed was never activated. For over three years, the model had nothing to look at. It only surfaced when the regulator asked why an apparently suspicious trade had gone unreported.

In another case, a coding mistake meant an alert required the firm to trade on the very day the price moved. In a liquid instrument, no problem. In a thinly traded bond? Potential insider dealing could pass entirely unseen. The model still produced good alerts, which gave the firm false comfort — the nastiest kind of quiet failure.

Why false positives are structural, not a bug

A spoofing alert is designed to catch traders placing orders they never intend to fill. But legitimate market-making and algorithmic execution also cancel orders constantly. A rule detects a pattern, not an intent. As one specialist puts it, the same cancellation pattern that suggests manipulation also appears in ordinary hedging.

Because missing abuse is far worse than over-alerting — a false negative can mean sanctions; a false positive costs analyst time — firms set thresholds conservatively. The queue grows. Then comes the tempting fix: raise the threshold, suppress the noisy rule, enjoy the shorter queue.

This is where things get dangerous. As one transaction-monitoring commentary puts it, a control isn't necessarily better just because it's quieter. False positives are visible because they land in the queue; false negatives are not. If a threshold change cuts alerts by a third, the obvious question is what stopped being flagged.

Calibration is a judgement, so document it

Good practice is unglamorous: diagnose before you recalibrate. Are analysts closing the same rule for the same reason? Is the noise concentrated in one customer segment? Has behaviour changed since the threshold was set? Then test alternatives against historical data — how many alerts would this have generated, which would have disappeared, would known suspicious cases still have been caught?

There is rarely one mathematically perfect number. What matters is that the reasoning is retrievable. "We raised the threshold because the team was overwhelmed" and "we tested three configurations and chose this one based on observed impact" are very different sentences to a regulator.

One practical wrinkle: the same price move means different things in different instruments. A 2% jump in a FTSE 100 stock is unremarkable; in a small-cap or a corporate bond it may deserve a look. Generic thresholds across asset classes produce either noise or blindness.

A gentle nudge about "all clear"

There is a subtler trap. If a system produces a reasonable number of decent alerts, everyone relaxes. Market Watch 79 describes firms concluding a model worked precisely because its output looked healthy — while a design flaw, years old, meant specific cases could never trigger.

Which suggests a useful habit: don't simply count alerts. Ask what isn't arriving. Is all the order and trade data actually being ingested — including orders that never execute, cancelled orders and amended ones? Data gaps are a recurring theme. Surveillance is only as good as the data underneath it.

Where machines can help — carefully

The FCA ran a 2024 TechSprint exploring AI and machine learning in surveillance. Teams presented approaches including isolation forests to reduce false positives, Bayesian networks for anomaly detection, and large language models to contextualise outliers. One idea that keeps recurring: score each rule-generated alert by likelihood, so human analysts spend their hours on the genuinely odd cases, with a full audit trail for anything closed automatically.

Another is behavioural profiling — judging a client's cancellation rate against their own history rather than a universal yardstick. An 80% cancellation rate is alarming for one client and entirely normal for an algorithm whose strategy is built on it.

Machines are not a shortcut past governance, though. The regulator is explicit that AI adoption needs oversight keeping pace. Testing needs to cover parameters, logic, code and data — and firms using vendor systems should understand how the alerts actually fire rather than trusting the brochure.

The takeaway for grown-ups

Surveillance tuning is a permanent balancing act, not a one-off project. Test formally and regularly. Watch the data, not just the alerts. Back up every calibration decision with evidence. And stay humble: the quietest system is not automatically the safest. It may simply have stopped listening.

Ovi
Ask OviThis entire article is in Ovi's context. Ask in any language.
I’m here with the complete article in front of me. What would you like me to explain?

Educational explanations only—not personalised trading instructions.

Research sources reviewed

  1. Market Watch 79 (fca.org.uk)
  2. Market Abuse Regulation (fca.org.uk)
  3. Market Abuse Surveillance TechSprint (fca.org.uk)
  4. 7 common market abuse scenarios monitored through trade surveillance (bloomberg.com)
  5. Fixing Trade Surveillance: Why 95% False Positive Rates Are a ... (insightcentric.com)
  6. Market abuse regulation: meeting FCA expectations (grantthornton.co.uk)
  7. False Positives in Transaction Monitoring: Why Calibration Matters (wearefortify.ai)
  8. Market abuse surveillance: How to get it right (hlc.com)
  9. Trade Surveillance Best Practices (eventus.com)
  10. Managed market abuse surveillance (ocorian.com)
  11. Navigating the FCA's SYSC rules (globalrelay.com)
  12. Governance of Market Abuse Surveillance Controls (afme.eu)

Comments

Comments are coming soon.