The Agentic Bank

VaR & Sensitivity Surveillance Agent

⬡ Quantile Surveils VaR, Greeks and back-testing for anomalies and explains the moves.
◆ Autonomous Specialist

Watches the daily risk numbers for anomalous moves … a VaR jump with no obvious cause, a back-testing exception, a sensitivity that drifted from the prior day … and writes the attribution. Runs the P&L-attribution pass overnight and surfaces the commentary by open.

Memory

Working Today's VaR / sensitivity vectors and the deltas vs. yesterday.
Episodic Prior back-testing exceptions and their explanations.
Semantic VaR methodology, risk-factor taxonomy, back-testing rules (Basel).
Store Results warehouse + file-based memory tool

Orchestration

MCP

Harness · Managed Agents … sandboxed code execution for risk-factor decomposition; structured note-taking across the daily surveillance run.

Tools

{ } Risk results warehouse API ›_ Risk-factor decomposition Code exec { } Market data feeds API ›_ Back-testing engine Code exec

Evals & guardrails

  • Explanation accuracy sampled by an agent-as-judge pass; drift detection on methodology.
  • Back-testing exceptions over a band require market-risk oversight-agent sign-off.
  • Reproducible, fully-traced calculations retained for model-risk audit.

Offline reflection

Offline consolidation of confirmed explanations builds a library of recurring VaR-move patterns, so the next anomaly is recognised, not re-derived from scratch.

Frontier edge

  • Causal attribution beyond decomposition: distinguishes a risk-factor shift from a data-glitch mark by testing counterfactual 'what if this input held' scenarios.
  • Continual learning: each validated explanation feeds an eval-gated update to the pattern library, so recurring moves stop needing a fresh derivation.
  • Long-horizon autonomy across a multi-hour overnight chain, surfacing a reviewed back-testing narrative by market open.

In numbers

ready at open
Morning commentary lead time
100%
Back-testing exceptions explained

Handoffs

Across ⇢ Model Risk Management for VaR-model validation

More on the Market Risk desk