Practitioner Perspective¶

Every derivatives desk operates in the gap between theory and reality. The risk-neutral formula $V_0 = \mathbb{E}^{\mathbb{Q}}[e^{-rT}X_T]$ assumes continuous trading, zero frictions, and known parameters---none of which hold. This section examines the compensating mechanisms that practitioners have built around the theoretical framework: calibration, model risk management, the gamma P&L decomposition, and the implied volatility surface.

Guiding principle

Practitioners do not trust models---they use them to structure risk and monitor deviations. The goal is not to find the "correct" model but to choose one that prices liquid instruments consistently, produces stable hedge ratios, and makes model risk transparent.

Scope

This section treats the pricing framework as an applied tool. It assumes familiarity with the core material in Physical vs Risk-Neutral World and Pricing vs Hedging. For the conditions under which the framework fails entirely, see When Measure Change Fails.

Calibration vs Estimation¶

Pricing models are fitted to market prices, not to history. A desk that estimates volatility from past returns and plugs it into the Black-Scholes formula will misprice every liquid option on the screen---and create apparent arbitrage against its own book.

Core idea

Calibration operates under the risk-neutral framework; estimation operates under the physical measure (see § Physical vs Risk-Neutral World).

The Two Approaches¶

	Estimation	Calibration
Measure	$\mathbb{P}$	$\mathbb{Q}$
Data source	Historical time series	Current option prices
Method	MLE, GMM, Bayesian	Least squares on prices
Output	Physical dynamics	Risk-neutral dynamics
Used for	Risk management, VaR	Derivative pricing

Why Calibration Dominates¶

Three reasons make calibration the standard for pricing:

Market consistency. A model that disagrees with observable option prices would create arbitrage against the desk's own book.
Information content. Implied volatilities encode the market's collective assessment of jump risk, stochastic volatility, and tail events that historical data may underrepresent.
Drift irrelevance. Estimating the physical drift is unnecessary for pricing (see § Risk Premium Decomposition).

Calibration does not determine the physical measure

A model calibrated to $\mathbb{Q}$ cannot be used for risk management tasks (VaR, expected shortfall) that require $\mathbb{P}$-measure forecasts.

The Calibration Problem¶

Given a model with parameter vector $\boldsymbol{\alpha}$ and $N$ liquid instruments with market prices $C_1^{\mathrm{mkt}}, \ldots, C_N^{\mathrm{mkt}}$, calibration solves

\[ \hat{\boldsymbol{\alpha}} = \arg\min_{\boldsymbol{\alpha}} \sum_{i=1}^{N} w_i\!\left(C_i^{\mathrm{model}}(\boldsymbol{\alpha}) - C_i^{\mathrm{mkt}}\right)^2 \]

where $w_i$ are weights (often inversely proportional to bid-ask spreads) and each model price is computed via risk-neutral valuation:

\[ C_i^{\mathrm{model}}(\boldsymbol{\alpha}) = \mathbb{E}^{\mathbb{Q}_{\boldsymbol{\alpha}}}\!\left[e^{-\int_0^{T_i} r_s\,ds}\,\Phi_i(S_{T_i})\right] \]

Non-uniqueness of calibration

In incomplete markets, different parameter vectors may fit market prices equally well yet produce different exotic prices. This is a manifestation of measure non-uniqueness; see When Measure Change Fails.

Model Risk¶

When two models calibrate perfectly to the same vanilla surface yet disagree on the price of an exotic, the difference is model risk. It is not a bug---it is an unavoidable consequence of market incompleteness.

Core idea

Different models calibrated to the same data select different risk-neutral measures and can disagree on exotic prices.

Sources¶

Model risk arises at three layers:

Specification risk. The choice of dynamics (Black-Scholes vs Heston vs local volatility) determines the set of available risk-neutral measures.
Calibration risk. Within a given model, different calibration sets or objective functions lead to different parameter estimates.
Extrapolation risk. Exotic prices require extrapolation beyond the liquid calibration instruments.

Quantification¶

A natural measure of model risk is the pricing interval:

\[ \text{Model risk} = \sup_{\mathcal{M}} V^{\mathcal{M}}(\Phi) - \inf_{\mathcal{M}} V^{\mathcal{M}}(\Phi) \]

where the supremum and infimum range over all models $\mathcal{M}$ that calibrate to the same liquid instruments.

Model risk in barrier option pricing

A down-and-out call calibrated to the same vanilla surface under three models:

Model	Barrier option price	Difference from BS
Black-Scholes (flat vol)	$4.82	baseline
Local volatility (Dupire)	$5.41	+12%
Stochastic volatility (Heston)	$5.18	+7%

Barrier options depend on the joint distribution of path and terminal value, which vanilla prices (marginal distributions) do not uniquely determine.

Hedging in Practice: The Gamma P&L¶

Perfect replication is a theoretical limit. In practice, discrete rebalancing and uncertain volatility produce a residual P&L that is driven entirely by gamma exposure.

Core idea

Discrete hedging generates P&L proportional to gamma times the gap between realized and implied variance.

Discrete Hedging Error¶

Between rebalance times $t_k$ and $t_{k+1}$, the P&L of a delta-hedged position is approximately

\[ \text{P\&L}_{k} \approx \frac{1}{2}\Gamma_{t_k}\,S_{t_k}^2\left[(\Delta W_k)^2 - \sigma^2\Delta t\right] \]

where $\Gamma_{t_k} = \partial^2 V / \partial S^2$ and $\Delta W_k = W_{t_{k+1}} - W_{t_k}$. Three consequences follow immediately:

P&L scales with gamma. Near-the-money, near-expiry positions have the largest hedging errors.
Realized vs implied variance drives the sign. The term $(\Delta W_k)^2 - \sigma^2\Delta t$ compares actual squared moves to the model's prediction.
The expectation is zero only when the model is correct. Under $\mathbb{P}$, the expected P&L per step vanishes only if $\sigma$ equals realized volatility.

Aggregate Gamma P&L¶

Summing over the option's life:

\[ \text{Total P\&L} \approx \frac{1}{2}\sum_{k=0}^{N-1}\Gamma_{t_k}\,S_{t_k}^2\left(\sigma_{\mathrm{realized},k}^2 - \sigma_{\mathrm{implied}}^2\right)\Delta t \]

Long gamma profits when realized volatility exceeds implied; short gamma profits in the opposite regime.

The trader's rule of thumb

"Buy options when you think realized vol will exceed implied; sell when you think the opposite." This connects the $\mathbb{P}$-measure forecast of future volatility to the $\mathbb{Q}$-measure implied volatility.

Transaction Costs and Hedging Frequency¶

Transaction costs create a tradeoff: more frequent rebalancing reduces hedging error but increases trading costs. For proportional costs at rate $\kappa$ per dollar traded:

\[ \text{Transaction costs} \sim \kappa\,\sigma\,S_0\,\sqrt{\frac{T}{\Delta t}} \]

\[ \text{Hedging error} \sim \Gamma\,\sigma^2\,S_0^2\,\sqrt{\Delta t} \]

The total cost is minimized at an optimal $\Delta t$ that depends on $\kappa$, $\sigma$, and $\Gamma$. Common strategies include:

Time-based rebalancing: hedge at fixed intervals (e.g., daily).
Threshold-based rebalancing: hedge when the delta deviation exceeds a tolerance band.
Utility-based hedging: minimize expected utility loss inclusive of transaction costs.

The Implied Volatility Surface¶

The Black-Scholes model predicts a flat implied volatility surface. Markets produce a rich two-dimensional surface $\sigma_{\mathrm{imp}}(K, T)$ that varies with both strike and maturity---and this surface is the market's risk-neutral distribution.

Core idea

The implied volatility surface encodes all risk-neutral marginal distributions via the Breeden-Litzenberger formula.

Empirical Features¶

Volatility skew: out-of-the-money puts carry higher implied volatility, reflecting crash risk and leverage effects.
Term structure: short-dated implied volatility spikes during stress and compresses in calm markets.
Surface dynamics: the surface moves stochastically, and no single parametric model captures its full dynamics.

Breeden-Litzenberger Formula¶

Each maturity slice of the surface defines a risk-neutral density:

\[ \frac{\partial^2 C}{\partial K^2}\bigg|_{K=k} = e^{-rT}\,q(k, T) \]

where $q(k, T)$ is the risk-neutral density of $S_T$ at level $k$. Calibrating a model to the surface is equivalent to fitting its risk-neutral marginal distributions to those implied by the market.

Calibration in Practice: Local Volatility¶

The Dupire local volatility model uses a state-dependent volatility $\sigma_{\mathrm{loc}}(S, t)$ chosen to reproduce all vanilla prices exactly:

\[ dS_t = rS_t\,dt + \sigma_{\mathrm{loc}}(S_t, t)\,S_t\,dW_t^{\mathbb{Q}} \]

Dupire's formula extracts $\sigma_{\mathrm{loc}}$ directly from the surface:

\[ \sigma_{\mathrm{loc}}^2(K, T) = \frac{\frac{\partial C}{\partial T} + rK\frac{\partial C}{\partial K}}{\frac{1}{2}K^2\frac{\partial^2 C}{\partial K^2}} \]

Local volatility provides a complete model (unique $\mathbb{Q}$) that fits all vanillas. However, its forward volatility dynamics are unrealistic, leading to poor performance for path-dependent and forward-starting options.

Theory vs Practice: Summary¶

Theoretical Assumption	Practical Reality
Continuous trading	Discrete rebalancing
Zero transaction costs	Bid-ask spreads, commissions, market impact
Known volatility $\sigma$	Stochastic, estimated or implied
Unique risk-neutral measure	Multiple models, multiple calibrations
Complete markets	Most markets are incomplete
Infinite liquidity	Limited depth, especially in stress

Despite these gaps, the framework provides a consistent pricing language, a starting point for hedging, and a benchmark against which model risk can be measured.

Trader translation

Price with implied vol ($\mathbb{Q}$).
Make money with realized vol ($\mathbb{P}$).

A practitioner's view

"All models are wrong, but some are useful." --- George E. P. Box

The risk-neutral framework is not a description of reality. It is a tool that translates derivative valuation into a tractable mathematical structure whose value lies in the discipline and consistency it brings to pricing and risk management.

Exercises¶

Exercise 1. A stock has physical drift $\mu = 0.15$, volatility $\sigma = 0.25$, and risk-free rate $r = 0.04$. A desk calibrates Black-Scholes to at-the-money options and obtains $\sigma_{\mathrm{imp}} = 0.22$. Explain why the calibrated volatility differs from the physical volatility. Which should be used for pricing a European call under $\mathbb{Q}$?

Solution to Exercise 1

The physical volatility $\sigma = 0.25$ is estimated from historical returns under $\mathbb{P}$. The implied volatility $\sigma_{\mathrm{imp}} = 0.22$ is extracted from option prices and reflects the risk-neutral distribution under $\mathbb{Q}$.

They differ because:

Different measures. Physical volatility is a $\mathbb{P}$-quantity; implied volatility is a $\mathbb{Q}$-quantity. In models richer than Black-Scholes (stochastic volatility, jumps), implied volatility is a nonlinear transformation of the risk-neutral dynamics and need not equal physical volatility.
Variance risk premium. Empirically, implied volatility tends to exceed realized volatility on average. The observation $\sigma_{\mathrm{imp}} < \sigma$ here is atypical but can occur in specific market conditions.

For pricing under $\mathbb{Q}$, the desk should use $\sigma_{\mathrm{imp}} = 0.22$. Using $\sigma = 0.25$ would produce a price inconsistent with the market, creating apparent arbitrage against the desk's book.

Exercise 2. Suppose two parameter vectors $\boldsymbol{\alpha}_1$ and $\boldsymbol{\alpha}_2$ both achieve the global minimum of the calibration objective for vanilla options. Explain why an exotic barrier option may still have different prices under $\boldsymbol{\alpha}_1$ and $\boldsymbol{\alpha}_2$, and relate this to market incompleteness.

Solution to Exercise 2

Both parameter vectors fit vanillas perfectly. However, vanilla prices determine only the marginal distributions of $S_T$ at each maturity (via Breeden-Litzenberger). They do not determine the path dynamics.

A barrier option's payoff depends on whether $S_t$ hits the barrier during the option's life---a path-dependent event governed by the joint distribution of $(S_t)_{0 \leq t \leq T}$, not just the terminal marginal. Two parameter vectors can produce identical marginals but different path dynamics (different local volatility surfaces, different spot-vol correlations, different jump structures).

This is market incompleteness: vanilla options do not span all contingent claims. Each $\boldsymbol{\alpha}_j$ implicitly selects a different risk-neutral measure from the family consistent with the vanilla surface, and these measures can disagree on path-dependent claims.

Exercise 3. A delta-hedged short call has $\Gamma = 0.04$, $S = 100$, $\Delta t = 1/252$, $(\Delta W)^2 = 0.006$, and $\sigma^2 \Delta t = 0.0004$. Compute the gamma P&L and determine the sign for the short call holder.

Solution to Exercise 3

Applying the gamma P&L formula:

\[ \text{P\&L} \approx \frac{1}{2}\Gamma\,S^2\left[(\Delta W)^2 - \sigma^2 \Delta t\right] \]

Substituting:

\[ \text{P\&L} \approx \frac{1}{2} \times 0.04 \times 10{,}000 \times (0.006 - 0.0004) = 200 \times 0.0056 = 1.12 \]

This gives $+\$1.12$ for a long gamma position. The short call has negative gamma, so the short call holder's P&L is $-\$1.12$. The position lost money because realized volatility far exceeded implied---the standard result for short gamma positions.

Exercise 4. Transaction costs scale as $\kappa\,\sigma\,S_0\,\sqrt{T/\Delta t}$ and hedging error scales as $\Gamma\,\sigma^2\,S_0^2\,\sqrt{\Delta t}$. Minimize the total cost with respect to $\Delta t$ to derive the optimal rebalancing interval.

Solution to Exercise 4

The total cost is

\[ C(\Delta t) = \kappa\,\sigma\,S_0\,\sqrt{\frac{T}{\Delta t}} + \Gamma\,\sigma^2\,S_0^2\,\sqrt{\Delta t} \]

Taking the derivative and setting it to zero:

\[ \frac{dC}{d(\Delta t)} = -\frac{1}{2}\kappa\,\sigma\,S_0\,\sqrt{T}\,(\Delta t)^{-3/2} + \frac{1}{2}\Gamma\,\sigma^2\,S_0^2\,(\Delta t)^{-1/2} = 0 \]

Multiplying both sides by $2(\Delta t)^{3/2}$:

\[ \Gamma\,\sigma^2\,S_0^2\,\Delta t = \kappa\,\sigma\,S_0\,\sqrt{T} \]

\[ \Delta t^* = \frac{\kappa\,\sqrt{T}}{\Gamma\,\sigma\,S_0} \]

The optimal interval increases with $\kappa$ (trade less when costs are high) and decreases with $\Gamma$, $\sigma$, and $S_0$ (trade more when hedging error is large).

Exercise 5. Using the Breeden-Litzenberger formula, explain why a symmetric implied volatility smile implies zero skewness in the risk-neutral distribution. What does the typical equity skew (higher implied vol for low strikes) say about the risk-neutral density?

Solution to Exercise 5

The Breeden-Litzenberger formula

\[ q(k, T) = e^{rT}\frac{\partial^2 C}{\partial K^2}\bigg|_{K=k} \]

extracts the risk-neutral density from the call price surface. Given a smooth implied volatility surface $\sigma_{\mathrm{imp}}(K, T)$, the call price $C(K,T) = \mathrm{BS}(S_0, K, T, r, \sigma_{\mathrm{imp}}(K,T))$ is twice differentiable in $K$, and $q(k,T)$ can be computed via the chain rule.

If the smile is symmetric around the at-the-money strike ($\sigma_{\mathrm{imp}}(K_0 + \delta) = \sigma_{\mathrm{imp}}(K_0 - \delta)$ for all $\delta$), the resulting density $q(k,T)$ is symmetric, giving zero skewness.

In equity markets the smile is asymmetric: out-of-the-money puts (low $K$) carry higher implied vol than out-of-the-money calls (high $K$). This produces a risk-neutral density with a fatter left tail---negative skewness---reflecting crash risk priced by the market.

Exercise 6. Suppose call prices are given by Black-Scholes with constant implied volatility $\sigma_0$. Show that Dupire's formula reduces to $\sigma_{\mathrm{loc}}(K, T) = \sigma_0$ for all $K$ and $T$.

Solution to Exercise 6

Under constant $\sigma_0$ the Black-Scholes call price is $C = S_0 N(d_1) - Ke^{-rT}N(d_2)$ with standard $d_1, d_2$. The required partial derivatives are:

\[ \frac{\partial C}{\partial T} = S_0\,n(d_1)\frac{\sigma_0}{2\sqrt{T}} + rKe^{-rT}N(d_2) \]

\[ \frac{\partial C}{\partial K} = -e^{-rT}N(d_2) \]

\[ \frac{\partial^2 C}{\partial K^2} = e^{-rT}\frac{n(d_2)}{K\sigma_0\sqrt{T}} \]

Numerator of Dupire's formula:

\[ \frac{\partial C}{\partial T} + rK\frac{\partial C}{\partial K} = S_0\,n(d_1)\frac{\sigma_0}{2\sqrt{T}} + rKe^{-rT}N(d_2) - rKe^{-rT}N(d_2) = S_0\,n(d_1)\frac{\sigma_0}{2\sqrt{T}} \]

Using the Black-Scholes identity $S_0\,n(d_1) = Ke^{-rT}n(d_2)$, the numerator becomes $Ke^{-rT}n(d_2)\frac{\sigma_0}{2\sqrt{T}}$.

Denominator:

\[ \frac{1}{2}K^2 \cdot e^{-rT}\frac{n(d_2)}{K\sigma_0\sqrt{T}} = \frac{Ke^{-rT}n(d_2)}{2\sigma_0\sqrt{T}} \]

Dividing:

\[ \sigma_{\mathrm{loc}}^2 = \frac{Ke^{-rT}n(d_2)\frac{\sigma_0}{2\sqrt{T}}}{\frac{Ke^{-rT}n(d_2)}{2\sigma_0\sqrt{T}}} = \sigma_0^2 \]

Therefore $\sigma_{\mathrm{loc}}(K,T) = \sigma_0$ for all $K, T$. A flat implied volatility surface produces a flat local volatility surface, confirming that Black-Scholes is the constant-local-volatility special case.

	Estimation	Calibration
Measure	\(\mathbb{P}\)	\(\mathbb{Q}\)
Data source	Historical time series	Current option prices
Method	MLE, GMM, Bayesian	Least squares on prices
Output	Physical dynamics	Risk-neutral dynamics
Used for	Risk management, VaR	Derivative pricing