What is model calibration in risk management and how do you avoid overfitting?

Model calibration is one of the most practically important topics in quantitative risk management. It's the process of selecting model parameters so that the model's outputs match observed market data as closely as possible — while remaining robust out of sample.

Calibration vs. Estimation:

Estimation looks backward; calibration looks at the present. For risk management, both are used — estimation for historical models, calibration for pricing models.

Calibration Example — Whitfield Derivatives:

Whitfield calibrates a local volatility model to match 30 listed option prices on an equity index. The objective function minimizes:

sum_{i=1}^{30} (Model_Price_i - Market_Price_i)^2

They have 8 model parameters. After optimization, the model matches all 30 prices within $0.02. Is this good? Not necessarily.

The Overfitting Problem:

If Whitfield had used 30 parameters for 30 prices, they'd get a perfect fit — but the model would have zero predictive power. Each parameter would memorize one data point rather than capturing the underlying volatility dynamics.

Signs of Overfitting:

1. In-sample fit is excellent but out-of-sample performance collapses — the model predicts last month perfectly but fails on new data
2. Parameters are unstable — recalibrating daily produces wildly different parameters
3. Parameters lack economic meaning — values that make no financial sense (e.g., negative volatility of volatility)
4. Excessive complexity — more parameters than justified by the data

How to Avoid Overfitting:

1. Parsimony: Use the simplest model that captures the essential features. The EWMA model (1 parameter) often outperforms complex models in VaR forecasting.
2. Out-of-sample testing: Calibrate on 80% of data, test on 20%. If performance degrades sharply, you're overfitting.
3. Regularization: Add a penalty term for parameter magnitude: minimize [Sum of squared errors + lambda x Sum of parameters^2].
4. Cross-validation: Rotate which data is in-sample vs. out-of-sample.
5. Economic constraints: Restrict parameters to economically sensible ranges.

Backtesting as the Final Check:

After calibration, backtest the model. For VaR at 99%, you expect ~2.5 exceptions per year (250 trading days x 1%). If you see 0 exceptions, the model is too conservative. If you see 10+, it's too aggressive. The Basel traffic light system formalizes this:

• Green: 0-4 exceptions
• Yellow: 5-9 exceptions
• Red: 10+ exceptions

For exam preparation on model risk topics, check our FRM question bank.