How an AI Agent Should Treat an Outlier in Market Data
Practical guidance for an AI trading agent to detect anomalous price points, verify data provenance, respect owner limits, and choose a safe response without
Produced with automation, then checked by deterministic quality rules and an independent source-grounded review before publication.
- 01Outliers should be flagged based on statistical deviation and freshness metadata.
- 02Verification of the data source and timestamp is required before any trade decision.
- 03Risk limits defined by the owner key must be consulted when an outlier suggests large exposure.
- 04An AI agent should default to a safe state‑pause or request human review‑when confidence is low.
- 05All actions taken on an outlier must be logged for later reconciliation and audit.
When market data shows a price that deviates sharply from recent values, an AI agent must first recognize it as an outlier, then verify its provenance, and finally decide whether to act, pause, or seek human input. This process protects the agent from acting on erroneous information while keeping risk controls intact.
What Defines an Outlier in Market Data?
An outlier is typically a data point that lies far outside the normal range of recent observations. Statistical methods such as z‑scores, interquartile ranges, or rolling volatility bands can quantify the deviation. Raw numbers alone are insufficient; the agent must also examine the data’s source, timestamp, and any freshness warnings supplied by the market feed.
For liquid instruments, a common rule of thumb is to treat any price that moves more than three standard deviations from a rolling mean as a candidate outlier. For less liquid assets, a wider band may be appropriate, but the principle remains the same: the deviation must be measurable and repeatable.
How Can an AI Agent Verify the Data Before Acting?
Verification involves three checks: source authentication, timestamp consistency, and coverage completeness. The agent should compare the reported exchange identifier with a whitelist of trusted venues and confirm that the timestamp is within an acceptable latency window. If any warning flags are present, the data must be treated as suspect.
- Cross‑check the price against at least one additional venue when possible.
- Reject data that lacks a valid timestamp or shows a freshness warning.
- Log the verification outcome for later reconciliation.
The verification step should also record the exact reason for any rejection. This log becomes essential when reconciling trades after a market disruption, because it provides a clear audit trail that distinguishes a deliberate pause from a system error.
Should the Agent Trade on an Outlier or Pause?
The safest default is to pause execution and either request human confirmation or wait for a second, corroborating data point. If the outlier passes verification and falls within owner‑defined risk limits, the agent may proceed, but only after applying conservative order sizing.
- 01Check owner‑signed limits for order size, daily notional, and daily loss before any trade.
- 02If limits would be exceeded, abort the trade and generate an alert.
- 03When limits are respected, place a modest order that can be quickly unwound if the market corrects.
A modest order size reduces exposure while still allowing the agent to capture a potential arbitrage opportunity. The order should be constructed so that it can be cancelled or reversed within a short time frame, typically a few seconds, to limit the impact of a rapid price correction.
What Are the Risks of Ignoring an Outlier?
Acting on a false price can lead to immediate loss of capital, unintended exposure, and downstream reconciliation headaches. Because market data may be delayed or corrupted, an unverified outlier can also cause the agent to breach policy limits, triggering emergency stop mechanisms that may not automatically close positions.
In addition, repeated exposure to erroneous data erodes confidence in the automated system and may require a manual review of the data‑feed contracts, increasing operational overhead.
How Does an Emergency Stop Interact With Outlier Handling?
If the agent detects a repeated pattern of outliers or fails verification repeatedly, it should invoke an emergency stop that revokes the calling key. This halts further automated activity but does not close existing positions or token allowances; those require separate owner review.
The emergency stop is a protective measure that preserves the integrity of the agent’s key hierarchy. By revoking the calling key, the system ensures that no further orders can be placed until the owner re‑authorizes activity after a thorough investigation.
A well‑designed AI agent treats uncertainty as a signal to pause, not to push forward.
Where Can I Learn More About Related Practices?
- How an AI Agent Should Handle Partial Fills
- How an AI Agent Can Verify Market Data Before Placing an Order
- Essential Risk Limits Every AI Trading Agent Should Enforce
Frequently asked questions
A simple z‑score using a rolling mean and standard deviation is easy to implement and works well for many liquid instruments.
No. Unknown sources lack the authentication needed for reliable decisions and should be rejected outright.
Ideally no. Cross‑checking with at least one other trusted venue reduces the chance of acting on a single‑point error.
Abort the trade, log the event, and generate an alert for the owner to review the limit configuration.
Regular reviews‑at least quarterly‑help ensure the statistical thresholds and verification steps remain appropriate as market conditions evolve.
Sources and verification
Product claims in this article were checked against these first-party references. Runtime status remains authoritative for current availability.
- Felix documentationfirst party
- Felix machine referencefirst party
Build with Felix now.
Felix infrastructure is live through MCP and the API. The full trading app launches September 17.
An AI trading agent must handle order cancellations and replacements with clear controls, error handling, and verification. This guide explains the safe workflow, required checks, and how to manage uncertainty.
A robust preflight verification protects capital and ensures compliance. This guide outlines the critical data, risk, and authorization checks every AI trading order needs.