V3Qtor AG/Financial services & insurance/Risk, markets and reporting
Model Calibration Desk
Complex distributions fitted and refitted as the book changes; multivariate regression, causal impact and Monte Carlo on governed data; calibration drift reported before the regulator asks.
Weeklycalibration, from quarterly
- Practice
- Decision & Risk Sciences
- Sector
- Financial services & insurance, risk, markets and reporting
- Systems involved
- AI Governance for the model registry
- Where the number comes from
- The number is an engagement target1.
What it changes in the P&L
Capital held against model risk.
How the first weeks go
- Fit the data you haveDistributions, regressions, time series: we start with what the system records and say where the data is too thin to decide.
- Put a confidence interval on the decisionA rate, a risk score, a treatment match: each comes with its uncertainty and with the rule that says when a person decides instead.
- Calibrate on a rhythmBack-tests run weekly, drift is reported before anyone asks, and the model version is logged with every output.
In an engagement
An asset manager's risk function moved from quarterly to weekly calibration, with distribution fitting and VAR refitted as the book changed and drift reported before the regulator asked1.
What stays
Model Calibration Desk stays in the book after the engagement. Model and claims reviews run from its numbers, and its evidence is what the regulator sees. Hand-over comes when the number has held for a quarter.
Talk to us about Model Calibration Desk.
Tell us where it would run and what it would have to change. We will come back with what we would look at first, what the first quarter produces, and what it would take.
Notes and sources
- 1Illustrative. Anonymised engagement scenarios; figures indicate modeled or targeted impact and are not audited outcomes.