Data stack (trading agents)
Pipeline that ingests feeds, normalizes events, scores quality, computes point-in-time features and research context, and serves state to policy and risk.
Related: Read more →
Reference
AI trading glossary
Canonical, quotable definitions used across Signal Desk research. Prefer these wordings when citing concepts from our articles.
Data quality flags
Structured feed status—typically healthy, delayed, or dead—based on latency, gaps, sequence breaks, and heartbeats so agents can fail closed.
Related: Read more →
Point-in-time features
Feature values computed using only information available at or before a decision timestamp—required for honest training and evaluation.
Related: Read more →
Online feature store
Low-latency store of features served to live policies, aligned with historical definitions to avoid train/serve skew.
Related: Read more →
AI trading agent
Software that perceives market and contextual data, reasons under uncertainty, and acts or proposes trades with a degree of autonomy under explicit risk constraints.
Related: Read more →
Trading bot
An always-on process that maps market inputs to orders using logic that is mostly fixed at design time (rules, grids, thresholds), even if it consumes ML scores.
Related: Read more →
Decision graph
How much of a trading system’s control flow is fixed at design time versus adapted at runtime (rules → scores → learned policy → planner/tools → multi-agent).
Related: Read more →
Risk gate
An independent service or process that validates every order intent against limits (position, loss, allowlists, venue health) and can approve, resize, reject, flatten, or halt—outside the model’s control.
Related: Read more →
Kill switch
A fast control that cancels open orders and blocks new intents. It must run independently of the agent process, be tested, and trigger on loss, errors, data staleness, or human command.
Related: Read more →
Controlled autonomy
Design principle that autonomy is allowed only inside explicit capital, venue, and operational constraints; models propose, risk and ops dispose.
Related: Read more →
Hot path
Low-latency decision loop (milliseconds to seconds) for quoting, inventory skew, or cancel/replace—usually deterministic, not LLM-driven.
Related: Read more →
Warm / cold path
Slower decision loops (seconds to hours) suitable for research agents, playbook selection, rebalancing, and LLM tool use.
Related: Read more →
Shadow mode
Running the agent on live data while logging decisions without sending real orders (or without capital impact), to validate data, latency, and behavior.
Related: Read more →
Paper trading
Simulated fills on live or recorded markets under the same policy and risk stack, used as a promotion gate before limited live capital.
Related: Read more →
Walk-forward validation
Rolling train/validate/test windows over time with purge/embargo around label horizons, reporting outcome distributions across windows rather than one lucky path.
Related: Read more →
Lookahead bias / leakage
Using future information in features, labels, normalization, universes, or LLM context during historical evaluation, which inflates offline performance.
Related: Read more →
Evaluation ladder (Signal Desk)
Promotion path: unit tests → component evals → cost-aware simulation → walk-forward → paper/shadow → limited live → full live only after stability gates.
Related: Read more →
LLM / tool-using agent
An agent that uses a language model to plan, extract events, call tools, and emit structured intents—strong for research and narrative risk, weak as a pure price oracle.
Related: Read more →
ML-enhanced signal bot
A system where models score signals or features but execution and sizing remain largely rule-bound; adaptive scoring without full agentic control flow.
Related: Read more →
Reinforcement / policy agent
A learned policy mapping sequential market/state features to actions (including hold/cancel) under costs and constraints, often trained offline and frozen live.
Related: Read more →
Multi-agent trading system
Specialized agents (e.g., research, risk, execution) that coordinate under a supervisor or shared protocols; a topology that still requires a hard risk boundary.
Related: Read more →
Market-making / inventory agent
A mandate-first hybrid that manages quotes and inventory under adverse selection, fees, and latency—success measured in inventory and survival, not narrative quality.
Related: Read more →
On-chain execution agent
An agent that submits on-chain actions (swaps, LP, etc.) under custody, gas, and MEV constraints; irreversibility demands least-privilege keys and tight spend limits.
Related: Read more →
Order intent schema
Structured fields for a proposed trade (symbol, side, size, urgency, reason codes, confidence, evidence IDs) that risk can validate before execution.
Related: Read more →
Position & notional limits
Hard caps on size, gross/net exposure, leverage, and concentration that risk services enforce on every intent.
Related: Read more →
Drawdown / loss guard
Rules that reduce size or halt trading when intraday or multi-day losses breach thresholds.
Related: Read more →
Event extraction (LLM)
Turning messy text into typed market events (unlocks, exploits, listings) with citations, then validating before features or risk flags update.
Related: Read more →
Prompt injection (trading agents)
Malicious content in web/social tools that attempts to manipulate a tool-using agent; mitigated by allowlists, schemas, least privilege, and human escalation.
Related: Read more →
Agent ablation test
Comparing the full agent to variants without LLM tools, memory, or other modules to prove the “AI” component changes outcomes.
Related: Read more →