Identification · event data · rival explanations
6Z Algorithmic Behavior: Test the Book, Not an Imagined Actor
A fast cancellation, a refill at the same price and a sudden move can all be generated by automated trading. They can also reflect manual orders, several unrelated firms, a feed artifact, a risk limit or new public information. Price and candles do not identify the actor. A defensible 6Z study starts with observable messages and asks which explanations the data can actually reject.
- Actor identity
- Usually unavailable
- Required feed
- Sequenced events
- Price alone
- Insufficient
- Published result
- None
The identification problem
A Pattern Is Not a Participant
Anonymous central-limit-order-book data can describe the sequence of public book events. It generally cannot establish whether one participant or many participants produced them, whether an order was generated manually or automatically, or why it was canceled. Even participant-level regulatory data would require context and intent analysis. Treat “algorithmic” as a hypothesis class, not a chart label.
Observed
Messages and trades
Timestamped additions, modifications, cancellations, executions, quantities, prices and sequence numbers.
Constructed
Book state
Spread, displayed depth, queue position where supported, imbalance and replenishment after each valid event.
Inferred
Behavioral mechanism
Liquidity provision, urgency, inventory response or execution scheduling—each conditional on rival tests.
Not inferred
Identity and intent
No claim that a named actor, “institution,” bot or manipulator caused a pattern.
Synchronize before measuring
Build an Auditable Event Stream
Use the dated 6Z contract, not an undocumented continuous symbol. Preserve exchange timestamps, sequence numbers, trade corrections and order-event identifiers where the licensed dataset supplies them. Reconstruct the book deterministically and fail the affected interval when a sequence gap cannot be repaired.
| Record | Use | Failure condition |
|---|---|---|
| Order events | Add, modify, cancel and execution sequence | Aggregated bars cannot substitute |
| Book snapshots | Validate reconstructed depth at checkpoints | Snapshot cadence hides intervening queue events |
| Trades | Aggressor rule, size and price response | Unresolved corrections or duplicate prints |
| Contract state | Expiry, roll, session, pause and holiday tags | Unlabeled contract migration |
| Official events | SARB and U.S. release-relative windows | Scheduled shocks mixed into ordinary periods |
CME DataMine separately describes Market by Order and PCAP products. Dataset entitlement and fields must be verified before claiming queue-level behavior; no market-data purchase or original 6Z study was performed for this article.
Candidate measures, not magic thresholds
Measure Cancellation, Replenishment and Response
Register definitions before inspecting outcomes. Report distributions by time and state rather than compressing the book into one average. Thresholds must be learned only in the training sample and carried unchanged into validation.
| Candidate measure | Definition | What it cannot establish |
|---|---|---|
| Cancel-to-add ratio | Canceled displayed quantity divided by added quantity within a registered window | Whether cancellation was strategic, automated or deceptive |
| Resting lifetime | Time from accepted add to execution, modification or cancellation | Intent; latency and queue rules affect exposure |
| Replenishment latency | Time until displayed size returns near a depleted level | That the same participant refilled it |
| Trade response | Midquote change over fixed event-time and clock-time horizons | Permanent impact without longer-horizon controls |
| Order-flow imbalance | Registered signed event or trade measure | A guaranteed next-price direction |
valid messages
→
book state
→
registered event
→
response + uncertainty
Every pattern needs competition
Challenge the Automated-Liquidity Story
A result is useful only if it survives plausible alternatives. Match or stratify by spread, depth, volatility state, time of day, order size, event proximity and contract lifecycle. Repeat the test with different but preregistered event definitions and include null intervals selected without looking at future returns.
Information arrival
- Official announcement window
- Broad USD or related-market shock
- Response begins before the candidate book event
Mechanical explanation
- Roll migration or expiry proximity
- Session boundary, halt or feed recovery
- Tick-size and queue effects
Sampling explanation
- One volatile week dominates
- Multiple thresholds were tried
- Effect disappears after costs or in holdout
A legal and evidentiary boundary
Cancellation Data Alone Cannot Prove Spoofing
The CFTC’s interpretive guidance explains spoofing in terms of bidding or offering with intent to cancel before execution and evaluates market context, patterns, fills and other facts. A public book observer does not possess that complete evidentiary record. High cancellation, fleeting size or layering can be a surveillance lead; none is a verdict.
Write “displayed quantity was canceled within the registered interval” or “the pattern met the study’s alert definition.” Do not write “algorithms manipulated 6Z” unless an authoritative enforcement finding supports that exact claim.
The research output is a log
Record What Would Make the Claim Fail
For each hypothesis, store its registration time, data version, eligible contracts, event definition, rival explanations, statistical estimator, uncertainty interval, cost model, holdout dates and disposition. Preserve negative and null findings. A result is retired when it depends on a small set of dates, fails a neighboring threshold, reverses out of sample or cannot clear conservative execution costs.
Retain
Stable descriptive pattern
Effect survives registered controls and holdout, with uncertainty reported. It still does not identify an actor.
Narrow
State-dependent evidence
Claim applies only to a declared event, size, book state and contract lifecycle.
Reject
Rival explanation wins
Timing, roll, volatility or data quality explains the apparent behavior.
Unscorable
Required fields absent
Bars, partial depth or broken sequences cannot support the intended claim.
Worked protocol example
Test apparent replenishment without naming the replenisher
Suppose displayed offer quantity is depleted by trades and similar quantity appears at the same price within a short interval. Register the depletion threshold, replenishment window, minimum displayed quantity and price-tolerance rule before looking at subsequent returns. For every candidate event, construct matched controls from the same dated contract, time bin, spread band, depth band and volatility state. Compare the distribution of replenishment latency and later midquote response, with confidence intervals and the number of independent trading days.
Then attack the interpretation. Repeat after excluding official release windows, roll migration, feed-recovery periods and intervals with unresolved sequence gaps. Vary the threshold only through a preregistered sensitivity grid and count every variation in the research family. Check whether the result survives when events from the same burst are collapsed into one episode. A pattern that disappears after these controls is rejected. A pattern that remains is described as recurrent displayed replenishment under the tested conditions; it still does not establish that one algorithm, one firm or one intent produced it.
No original result is reported. This page publishes no 6Z actor classification, profitable signal or spoofing finding; it is a reproducible protocol and evidence boundary.
Sources, methods and editorial disclosure — reviewed August 25, 2026
- CME Rulebook index and Chapter 259 for 6Z’s ZAR/USD contract identity.
- CME DataMine for official historical-data product categories.
- BIS Markets Committee, FX execution algorithms and market functioning for market-wide execution-algorithm context.
- CFTC anti-disruptive-practices interpretive guidance for the intent and context boundary around spoofing.
Sources were reviewed August 25, 2026. This unsponsored article distinguishes exchange facts, proposed measures, inference and regulatory findings. It reports no original empirical result.