Calibration Stability And Regularization

Hand a flexible enough model a list of quoted option prices and ask it to match them, and it will — a model with enough free parameters can thread a curve through any finite set of points, the same way a high-degree polynomial bends through any scatter of dots. The trouble shows up the next day. One quote moves by a cent, you recalibrate, and the "best fit" parameters jump violently, even though nothing about the market actually changed. A model whose inputs swing wildly on noise is not a model of the market; it is an expensive way of memorising yesterday's data.

The overly flexible tailor

Imagine a tailor who refits your entire suit from scratch every time you gain or lose an ounce, rather than making a small adjustment. Each suit fits perfectly on the day it's made, but the tailor's method is unstable — a trivial weight change produces a completely different cut. A sensible tailor accepts a suit slightly loose in exchange for a method that responds smoothly to small changes. Regularization is that sensible tailor's rule, written as a penalty added to the calibration.

The penalized objective

Ordinary calibration finds parameters $\theta$ that minimise the gap between model prices $V_{\text{model}}(\theta)$ and market prices $V_{\text{mkt}}$:

$$\hat\theta = \arg\min_{\theta} \; \sum_i \big(V_{\text{model}}^i(\theta) - V_{\text{mkt}}^i\big)^2 .$$

In plain English: pick the parameters that make the model's prices as close as possible to what the market actually quotes. A regularized calibration adds a second term that punishes the parameters for moving too far from a stable reference, usually yesterday's calibrated values $\theta_0$:

$$\hat\theta = \arg\min_{\theta} \; \sum_i \big(V_{\text{model}}^i(\theta) - V_{\text{mkt}}^i\big)^2 \;+\; \lambda \, \lVert \theta - \theta_0 \rVert^2 .$$

In plain English: don't just match the quotes, match them while staying close to where you already were. The constant $\lambda$ (lambda) is a dial: at $\lambda = 0$ you get the pure, unstable best fit; as $\lambda$ grows, the model cares less about the last cent of price and more about not jumping around, until at very large $\lambda$ the parameters barely move regardless of the market.

Worked example 1: a one-parameter surface, two quotes

A toy model prices two options with a single volatility parameter $\theta$: $V_1(\theta) = 5 + 2\theta$ and $V_2(\theta) = 3 + 4\theta$. Market quotes today are $V_1 = 5.20$, $V_2 = 3.35$.

Unregularized fit minimises $(5+2\theta-5.20)^2 + (3+4\theta-3.35)^2$. Taking the derivative and setting it to zero: $4(2\theta - 0.20) + 8(4\theta - 0.35) = 0 \Rightarrow 8\theta - 0.80 + 32\theta - 2.80 = 0 \Rightarrow 40\theta = 3.60 \Rightarrow \theta = 0.090$.

Now suppose $V_2$'s quote ticks up one cent, to $3.36$. Redo it: $8\theta - 0.80 + 32\theta - 2.88 = 0 \Rightarrow 40\theta = 3.68 \Rightarrow \theta = 0.092$. A one-cent quote move shifted $\theta$ by $0.002$ — small here, but with more parameters and correlated quotes this compounds and can move $\theta$ far more.

Regularized fit with $\theta_0 = 0.085$ (yesterday's value) and $\lambda = 20$: minimise $(5+2\theta-5.20)^2 + (3+4\theta-3.35)^2 + 20(\theta-0.085)^2$. The derivative condition becomes $40\theta - 3.60 + 40(\theta - 0.085) = 0 \Rightarrow 80\theta = 3.60 + 3.40 = 7.00 \Rightarrow \theta = 0.0875$, pulled noticeably toward $\theta_0$ and away from chasing the last cent of fit.

Worked example 2: how much fit you give up

Using $\theta = 0.0875$, model prices are $V_1 = 5 + 2(0.0875) = 5.175$ (versus quoted $5.20$, off by $2.5$ cents) and $V_2 = 3 + 4(0.0875) = 3.35$ (exact). Squared pricing error is $(0.025)^2 = 0.000625$, tiny in dollar terms, versus zero error for the unregularized fit. That quarter-cent of mispricing is the entire cost of a model whose parameters no longer whip around on noise — a trade every serious calibration desk takes willingly.

What this means in practice

Every stochastic-vol or local-vol calibration run in production adds some form of this penalty — often a smoothness penalty on the local-vol surface itself, or a proximity penalty to yesterday's parameters as here — because a desk needs Greeks that are stable day to day, not just a fit that's perfect today. An unregularized model produces phantom hedging trades every morning as parameters swing on noise, costing real fees for zero economic reason.