Skip to main content

Risk Management

The platform is designed and tested so that every execution-capable bot and strategy path (see Trading Bots and ML Models & Strategy Layer) passes through the shared risk gate before an order can be submitted.

Core protections​

  • Drawdown limits and consecutive-loss breakers — a losing streak or drawdown threshold halts new entries for the affected strategy rather than relying on a human to notice and intervene.
  • Volatility gates and leverage caps — position sizing and leverage are constrained by current volatility regime (fed by the Regime analytics stream), not fixed at signup.
  • Correlation-aware position sizing — the risk layer accounts for cross-position and cross-asset correlation, not just single-position risk.
  • Cascade/liquidation risk blocking — the dedicated Cascade Predictor model (see ML Models) can veto entries during elevated liquidation-cascade risk.

Forward-paper calibration​

Strategy parameters are not fixed at deployment. The calibration layer tracks edge decay against forward-paper performance telemetry, runs A/B comparisons between parameter sets, and gates promotion of a new configuration behind a reconciliation check (see Metrics Methodology for the paper-vs-live-capital distinction). This is the same layer exposed read-only through the MCP calibration_* tools for observability — see Catalog Overview.

Recovery posture​

Risk-gate trips are designed and tested to fail toward inaction (halt new entries) rather than toward uncontrolled exposure — a detection or feed failure is intended to narrow what the system will do next, not default to "keep trading as if nothing happened."