“The model with the smallest calibration error is the best model.”
Not quite.Fit is one criterion; stability, identifiability, arbitrage behavior, interpretability, and intended use also matter.
Architecture & Systems
Model Calibration and Validation
Calibration fits model parameters to market data; validation checks whether the model is fit for purpose.
Mental model
Calibration fits parameters to observations, while validation tests whether the fitted model is useful for its purpose.
Calibration chooses parameters to minimise mispricing against a set of liquid instruments, subject to constraints.
Good calibration tracks quality metrics such as error distributions, stability over time, and parameter plausibility.
Explore the mechanism
Does the fitted parameter set reproduce calibration targets, and is the model reliable for the decision it will support?
This smile shows market implied vols as points and a polynomial model calibrated to part of them. Calibration error is measured on the calibration set, validation error on held-out points. Use the controls to change model flexibility and noise and see how underfit, good fit, and overfit behave.
Degree: 2 · RMSE(cal) ≈ 0.0062 · RMSE(val) ≈ 0.0141
Numbers
Degree: 2
RMSE (calibration) ≈ 0.0062 · RMSE (validation) ≈ 0.0141
Val / Cal RMSE ratio ≈ 2.27
Interpretation
Calibration is about matching the data you chose. Validation asks whether, given noise and structural limits, the model stays sensible on data you did not directly fit.
Underfit models leave systematic structure in the residuals. Overfit models drive calibration error down but see validation error explode as they chase noise. In a library, calibration and validation should produce a versioned artefact with both error metrics and rationale, not just a parameter vector.
UNDERSTAND
What to notice
Validation assesses whether the model is structurally appropriate for its use-case, beyond just calibration fit.
In a central library, calibration and validation produce versioned artefacts and documentation, not just numbers.
Where it appears
- fitting volatility and credit models
- approving pricing models for production
- monitoring model performance and stability
Where the intuition breaks
What this view leaves out
- A close calibration fit does not establish out-of-sample validity, parameter stability, or suitability for every product and risk measure.
Common mistakes
Where this concept sits
Follow the strongest prerequisites, applications, extensions, and contrasts.
Connections for Model Calibration and Validation
Model Calibration and ValidationYOU ARE HERE
Built onValidation Theory: Backtest Power and Multiple TestingValidation theory provides statistical discipline for assessing whether calibrated models actually perform out of sample.
UsesImplied Volatility Smile and SurfaceVolatility surfaces are common calibration targets and diagnostics for option-pricing models.
Validated byMonte Carlo Model ValidationMonte Carlo convergence and statistical diagnostics test the numerical evidence used in the broader model validation process.
ValidatesStochastic Volatility (Heston) and Characteristic FunctionsCalibration fits Heston parameters to market option data, while validation checks whether the fitted model reproduces relevant prices and dynamics.
SupportsMarket and Model VersioningCalibrated parameters and validation status should be tied to a specific model version.
ValidatesMonte Carlo PricingA Monte Carlo implementation should be calibrated where appropriate and validated against known prices, limits or benchmarks.
Built on · Validation Theory: Backtest Power and Multiple Testing
Validation theory provides statistical discipline for assessing whether calibrated models actually perform out of sample.
Explore concept →