What is toxic flow, and how do brokers detect it with AI?
Toxic flow is order flow that consistently generates adverse markouts against the counterparty — a client whose trades regularly move the market against the broker or the LP. Modern brokers detect it with a firm-specific model that scores every order 0–100 in real time, using client history, instrument context, session state and correlated flow, then feeds that score into A/B routing decisions with a dealer-approvable UI.
Toxic flow is order flow that consistently generates adverse markouts against the counterparty — a client whose trades regularly move the market against the broker or the LP. Modern brokers detect it with a firm-specific model that scores every order 0–100 in real time, using client history, instrument context, session state and correlated flow, then feeds that score into A/B routing decisions with a dealer-approvable UI.
Toxicity is not fraud — it is signal
Toxic flow is not a moral category. A trader whose orders consistently move the market against the counterparty is producing signal — often because they are informed, sometimes because they are exploiting a latency edge, occasionally because they are gaming a specific instrument. Whatever the cause, the response is not to block the client; it is to route their flow appropriately.
The modern detection stack
The pattern that works: every order arriving in the book is scored 0–100 in real time by a firm-specific model that reads client history, symbol context, session state and correlated flow. The score is not a black-box number — it comes with a reason string the dealer can read ('client has 87% adverse markout on EUR/USD spikes; correlated to two other flagged accounts in the last hour').
The score feeds the routing decision: A-book, B-book, hybrid, or a specific LP mix. The decision is dealer-approvable, versioned and audit-logged. The model updates on each session's evidence, not next quarter's.
Related questions
Isn't toxic-flow detection just a rule set?
The naive version is. Modern implementations use a firm-specific model that captures nonlinear interactions rules cannot — client history × symbol × session state × correlated flow all combined.
Can this be built on a bridge alone?
The data lives in the bridge. The model does not. Building it requires a firm-specific stack on top of PrimeXM/oneZero/Centroid data.
Does this replace the dealer?
No. The dealer approves. The model produces speed and accuracy of input the dealer cannot sustain manually.
What's the compliance posture?
Stronger than manual — every routing decision has a reason and a version, audit-logged.
Price?
Ships inside the Dealing Desk AI productised build (£120k / 8 weeks).
Toxic-flow scoring is the single highest-EV AI system on a dealing desk. The firms that shipped it in 2025 opened material gaps in per-trader economics that competitors are only now noticing.
The £15k AI Diagnostic sizes the specific toxic-flow lift for your book.