How ModelQuant Quantifies GenAI Risk

A deterministic Monte Carlo approach to measure GenAI Value-at-Risk and Expected Shortfall in euros.

1

Define Failure Modes

Input your GenAI model's failure modes: name, probability of occurrence, impact in euros, and pairwise correlation with other failure modes. ModelQuant supports any number of failure modes.

2

Set Confidence Level

Choose 95% or 99% confidence level. 95% is suitable for standard reporting; 99% aligns with regulatory capital calculations under SR 26-2.

3

Monte Carlo Simulation

ModelQuant runs 10,000 correlated draws through the Cholesky-decomposed failure tree. Each draw triggers failure events proportionally to their probabilities and correlation structure, with lognormal severity around the specified impact.

4

VaR & ES Calculation

The simulation results are sorted to produce the loss distribution. Value-at-Risk (VaR) is the loss at the specified percentile. Expected Shortfall (ES) is the average loss beyond VaR — capturing tail risk.

5

Component Breakdown

Each failure mode's contribution to total VaR and ES is decomposed. Heatmap data points and histograms are provided for visualization.

API Reference

Integrate ModelQuant directly into your risk infrastructure.

POST /api/assess

{
  "modelName": "GPT-4o",
  "confidenceLevel": 0.99,
  "failureModes": [
    {
      "name": "Hallucination",
      "probability": 0.08,
      "impactEUR": 250000,
      "correlation": 0.3
    }
  ]
}

Response

{
  "VaR": 425000.00,
  "ES": 680000.00,
  "componentBreakdown": [...],
  "heatmapData": [...],
  "histogram": [...],
  "simulationMetadata": {...}
}

Read VaR Insights (RSS) Try Demo