Quick answer: market-surveillance systems do not reliably infer manipulation from one order, one cancellation or one price move. They reconstruct sequences of orders, modifications, cancellations and executions across time, accounts, products and venues, then look for patterns inconsistent with bona fide trading. Rules can detect known structures; statistical and machine-learning models can rank unusual behaviour; graph models can reveal coordinated accounts or cross-product relationships. The output should be an alert for investigation, not an automatic finding of intent. Manipulation is a legal and evidentiary conclusion that requires context beyond the pattern score.

A surveillance alert says “this sequence deserves evidence.” It does not say “this trader is guilty.”

Safety boundary: this article explains defensive compliance and detection architecture. It intentionally avoids institution-specific surveillance thresholds, tuning values, evasion methods or instructions for disguising manipulative trading.

Page role: market-integrity surveillance, not AML transaction monitoring

Bukit Timah Tutor already has How Bank Transaction-Monitoring Algorithms Detect Money Laundering. That article focuses on customer money movements, AML context, peer groups and transaction relationships.

This page owns a different job: how trading firms and banks surveil order-entry and trading behaviour for possible market manipulation.

FINRA’s 2026 Regulatory Oversight Report identifies surveillance for potential layering, spoofing, wash trades, prearranged trades, marking the close and other manipulative schemes as a continuing supervisory obligation. It also stresses that surveillance parameters should fit the firm’s products and business and be reviewed as markets change. See FINRA — Manipulative Trading.

1. The raw material is an event stream

An electronic order book generates a sequence of timestamped events:

new order;
order modification;
partial execution;
full execution;
cancellation;
trade report;
quote update;
auction or session-state changes.

A surveillance engine first converts that stream into a coherent state history. If an order ID, timestamp, account identifier or parent-child linkage is wrong, the model can invent suspicious patterns that never happened or miss real ones that were split across systems.

This makes market surveillance a data-engineering problem before it becomes a machine-learning problem.

2. Reconstruct the limit order book through time

At each instant t, the order book contains bids and offers at different prices and quantities. A simplified state can be represented as:

B(t) = {bid prices/quantities, ask prices/quantities, recent trades, participant-linked events}.

The surveillance system is interested not only in the state but in the transition from B(t) to B(t+1). Did displayed depth appear and disappear? Did a trader execute on one side while repeatedly placing and cancelling orders on the other? Did activity move the visible book around a benchmark or close?

The sequence is often more informative than any single snapshot.

3. Spoofing is an intent concept; algorithms observe behaviour

US commodity law defines spoofing as bidding or offering with the intent to cancel the bid or offer before execution. CFTC guidance also makes clear that legitimate good-faith cancellation is not automatically spoofing and that context, patterns and facts matter. See the CFTC Interpretive Guidance and Policy Statement on Disruptive Practices.

This creates an important mathematical boundary:

The surveillance model observes order behaviour. It does not directly observe mental state.

Therefore a model can identify a sequence consistent with a manipulation hypothesis, but investigators still need surrounding evidence before concluding intent.

4. Sequence features turn raw events into measurable behaviour

High-level defensive features can summarise how a participant interacts with the book without relying on one order. Examples include:

order lifetimes and modification patterns;
executed versus cancelled quantity;
side-switching around executions;
changes in displayed depth associated with the participant’s events;
distance of orders from the prevailing best prices;
concentration of events around auctions, closes or benchmarks;
repeated order-execution-cancellation sequences;
cross-product timing relationships.

The exact surveillance features and thresholds should remain governed and institution-specific. Publishing a universal “safe cancellation ratio” would be both mathematically unsound and contrary to the purpose of defensive surveillance.

5. Rules encode known hypotheses

A rule engine can formalise a known pattern:

Alert if a documented sequence of order-entry, cancellation, execution and price/depth response occurs within a governed context.

The purpose of a rule is not to label a crime. It is to preserve a testable hypothesis that investigators know how to interpret.

FINRA’s 2026 report warns against generic controls that are not tailored to product classes or sources of order flow, and against thresholds set too high or too low to identify meaningful activity. That is an algorithm-design lesson: rules need a defined operating domain.

6. Why one cancellation ratio is a bad detector

Legitimate market making can involve frequent cancellations because prices, inventory and queue position change continuously. A high cancellation rate by itself therefore has low evidentiary specificity.

A stronger detector asks how cancellations relate to:

executions on the opposite side;
changes in visible market depth;
price movement;
repetition across time;
participant inventory;
market volatility;
legitimate quoting obligations.

This is a general modelling principle: a feature becomes more informative when embedded in the mechanism it is supposed to represent.

7. Layering is relational and temporal

“Layering” is commonly used to describe patterns involving multiple orders at different price levels that create a potentially misleading appearance of market interest while related activity seeks execution elsewhere. The defensive modelling challenge is to reconstruct the relationship between those orders rather than treating each as independent.

A sequence model can represent:

state before participant activity → participant order cluster → market response → participant execution → cancellation/withdrawal of remaining cluster.

Repeated recurrence of the same economic sequence can be more informative than the absolute number of orders.

8. Graphs detect coordination that one-account surveillance misses

Some suspicious trading patterns are distributed across accounts, legal entities, products or venues. A graph can represent:

accounts as nodes;
common ownership/control as edges;
trading against each other as edges;
shared devices or authorised persons where lawfully available;
common issuers or related instruments;
time-correlated order sequences.

Graph features can then surface dense clusters, repeated pairings, unusual circularity or cross-account coordination for human review.

FINRA specifically highlights monitoring across customers, platforms and related products as an effective practice for detecting manipulative schemes that do not fit inside one account or one venue.

9. Cross-market surveillance is a data-joining problem

A participant can trade a stock, option, future, ETF or economically related instrument. If surveillance sees only one market, it may miss the economic purpose of the activity.

Cross-market surveillance requires:

consistent instrument mapping;
common clocks and timestamp quality;
account/entity resolution;
economic relationships between instruments;
event sequencing across venues.

That turns manipulation detection into a graph-and-time-alignment problem as much as a classification problem.

10. Anomaly scores rank attention, not guilt

Unsupervised models can identify behaviour that differs from a trader’s history or a peer group. Supervised models can learn from previously reviewed cases. A generic score can be written:

score = g(sequence features, book response, execution features, account graph, product context, market regime).

The score is often best interpreted as a triage variable: which events should investigators review first?

A score of 0.92 does not mean “92% probability of manipulation” unless the model has been explicitly calibrated and validated to support that probabilistic interpretation. Ranking and probability are different model outputs.

11. False positives are structurally unavoidable

Many behaviours associated with manipulation can occur legitimately:

rapid cancellations during volatile markets;
large displayed orders that are genuinely intended to trade;
cross-account trades generated by independent strategies;
closing-period activity driven by index rebalancing;
repeated order replacement caused by smart-order-routing logic.

A surveillance system therefore needs precision as well as recall. Too much sensitivity creates an alert flood that investigators cannot review well. Too little sensitivity misses meaningful patterns.

The correct objective is not “minimise alerts.” It is “route scarce investigative capacity toward the highest-value evidence while preserving the ability to detect material behaviour.”

12. The labels are selected and incomplete

Supervised machine learning faces a familiar but serious problem: historical labels come from cases that the old surveillance system chose to review. If the old rules never surfaced a new manipulation pattern, that behaviour may sit in the training data labelled implicitly as normal.

This is selection bias. The dataset contains a history of what the institution noticed, not the complete universe of manipulative and legitimate activity.

Useful controls include:

random sampling of non-alerted activity;
regulator/exchange cases used as external challenge examples;
synthetic or simulated pattern testing for system verification;
challenger rules based on different hypotheses;
periodic search for novel clusters outside existing alert taxonomies.

13. Market regime matters

Order-book behaviour changes with volatility, liquidity, auctions, news events, tick size and market structure. A cancellation pattern that is unusual in a calm large-cap stock may be ordinary during a volatile futures release.

A robust model therefore conditions on market state or compares like with like. Otherwise the detector can confuse volatile market with manipulative participant.

This is one reason FINRA emphasises tailoring surveillance parameters to product class and periodically reviewing them as the firm, customer base and market change.

14. Human review adds evidence the order book does not contain

Investigators can examine:

full order and execution history;
related accounts and products;
communications where lawfully available and relevant;
strategy explanations;
market events;
customer instructions;
previous alerts;
other regulatory or exchange information.

The model compresses a large event stream into a manageable case. The investigator reconstructs the economic meaning.

That division is not a weakness of automation. It is a recognition that the observed sequence and the legal question of intent live at different evidentiary layers.

15. Singapore market-integrity context

Singapore Exchange rules prohibit trading members from engaging in or knowingly assisting practices likely to create a false or misleading appearance of active trading or a false market. See the SGX rulebook section on Market Manipulation and False Market.

The public educational lesson is jurisdiction-neutral: surveillance logic should map to the actual legal obligations and venue rules that govern the product. A US futures spoofing rule, a securities-market false-trading rule and an exchange’s disruptive-practices rule can overlap conceptually without being identical legal tests.

16. Evidence polarity: what should reduce suspicion?

A defensive model should be designed to accept exculpatory evidence, not only accumulate red flags. Evidence that can weaken a manipulation hypothesis includes:

consistent bona fide execution attempts;
market-making obligations explaining quote churn;
strategy behaviour that remains symmetric across directions;
cancellations explained by broad market repricing rather than participant-created pressure;
independent accounts with no coordination evidence;
a pattern that disappears after correcting data timestamps or parent-child order links.

A surveillance system that cannot become less suspicious when contrary evidence arrives is not doing investigation support. It is doing confirmation bias at machine speed.

17. Counterexamples that break simplistic surveillance

“High cancellation means spoofing.” Counterexample: legitimate high-frequency market making in a fast-moving book.
“Large displayed order means deception.” Counterexample: a genuine institutional order that is partially filled before market conditions change.
“Anomaly means manipulation.” Counterexample: a new but legitimate strategy outside the training distribution.
“No alert means normal.” Counterexample: a new cross-account pattern invisible to single-account rules.
“One venue is enough.” Counterexample: manipulation hypothesis depends on related activity in an option, future or correlated product elsewhere.

18. The defensive surveillance pipeline

Ingest complete order, modification, cancellation and execution events.
Synchronise clocks and reconstruct parent-child order identity.
Rebuild the order book and participant event sequence.
Map related accounts, products and venues where permitted.
Generate governed rule-based alerts for known hypotheses.
Calculate sequence, market-response and graph features.
Apply anomaly or classification models as triage tools.
Condition on product and market regime.
Prioritise alerts using materiality and evidence strength.
Investigate with broader contextual evidence.
Record reasons, not only final case labels.
Sample non-alerted activity to challenge blind spots.
Revalidate thresholds and models after market changes.
Escalate legal conclusions to authorised compliance/investigation functions.

19. Failure modes and weak links

Intent inference error. Behavioural pattern is treated as direct proof of mental state.
Threshold monoculture. One universal threshold is applied across products and regimes.
Order-ID fragmentation. Parent-child events are split and the reconstructed sequence is false.
Alert-volume optimisation. Teams reduce alerts without checking whether meaningful cases disappear.
Training-label bias. New model learns only the blind spots of the old model.
Cross-market blindness. Related product activity is not joined.
Human-review bottleneck. Good alerts age unreviewed behind noise.
Adversarial drift. Manipulative behaviour changes while rules remain static.

20. Diagnostics and falsifiers

Which alert type produces the highest volume and lowest investigative value?
How much does performance change by product and volatility regime?
Can independent reconstruction reproduce the same order sequence?
Do sampled non-alerted cases reveal patterns absent from current rules?
Which model features dominate alert ranking?
Does cross-market data materially change case outcomes?
How often do investigators overturn the model’s highest-risk cases?
What evidence would falsify the manipulation hypothesis for this alert?

Suppose someone claims, “The trader cancelled nearly all displayed orders, therefore spoofing is proven.” A falsifier is evidence that the orders were part of a bona fide quoting strategy, were exposed to genuine execution risk, cancellations tracked broad market repricing, and no related execution pattern supports an intent-to-deceive hypothesis. Cancellation frequency alone cannot carry the legal conclusion.

21. Verification and update triggers

reconcile surveillance feeds against exchange/broker records;
test timestamp and sequence reconstruction;
review alert performance by product and market state;
compare rule alerts with machine-learning challengers;
sample non-alerted activity;
recalibrate after tick-size, venue, product or order-type changes;
update typologies after regulator/exchange cases;
preserve investigator reasoning so future models learn from evidence rather than a bare binary label.

Connections across the finance-and-banking algorithms lane

AML transaction monitoring — similar alert-routing logic applied to money movement rather than order-book manipulation.
Model validation — needed because surveillance models operate under incomplete labels and drift.
Payment-system graphs — another example of event streams becoming networks and queues.
A Formula Can Be Correct and Still Be the Wrong Model — a warning against treating one statistical red flag as the whole legal/economic mechanism.

Research anchors

The deeper lesson

Market surveillance is sequence reasoning under incomplete evidence. The order book shows actions, not intent. Rules preserve known hypotheses. Sequence features reconstruct mechanism. Graphs reveal coordination. Machine learning ranks unusual behaviour. Human investigators return the alert to economic and legal context. The strongest surveillance system is therefore not the one that accuses fastest; it is the one that can explain why a pattern deserves attention, what evidence would clear it, and how the detection logic changes when the market itself changes.

Educational note: This article explains defensive market-integrity and computational concepts. It does not provide evasion techniques, trading advice, legal conclusions about any person or institution, or institution-specific surveillance thresholds.

Small Group Tutorials

How Bank Market-Surveillance Algorithms Detect Manipulation: Order Books, Spoofing, Layering, Sequence Features, Graphs, False Positives and Human Review