3.11.6 Advanced Extensions

Basic Prerequisites: 05 · Failure Modes and Interest Rate & Term Structure Models.

---

1. Intuition & Practical Objective

Two ways the discipline generalizes beyond a single equity smile, plus the desk-level workflow:

1. SVI - a stable, arbitrage-free way to parameterize the whole surface. Instead of inverting raw quotes into local vol (unstable, 05), Gatheral’s SVI (“stochastic-volatility-inspired”) fits a small parametric total-variance surface $w(k)=a+b\{\rho (k-m)+\sqrt{(k-m{)}^2+{\sigma }^2}\}$ per expiry. It is smooth and, fit carefully, has no strike arbitrage (butterfly $\ge 0$) - it is the standard intermediate step before any inversion.
2. Rates - calibration when the state space is a curve. In the LIBOR Market Model (LFM/BGM), the state is an entire forward-rate curve, so calibration is a curve-fitting problem with a structure: caplets fix the average (integrated) volatilities, swaptions fix the residual (correlation-shaped) parameters. Brigo–Mercurio’s cascade algorithm walks the swaption-volatility matrix one entry at a time, each step solving an algebraic quadratic in a single new volatility.
3. Desk practice - the workflow that keeps it all honest. Market data → conventions → smoothing → calibration → no-arbitrage check → residual report → recalibration.

The practical objective: see SVI fit a smile and pass a butterfly check, understand the LMM two-step/cascade philosophy, and get the concrete desk workflow.

---

2. Mathematical Ground Truth & Derivations

SVI (Gatheral eq 3.20). Per expiry, model total implied variance as

$$w(k)=a+b(\rho (k-m)+\sqrt{(k-m{)}^2+{\sigma }^2}),k=\ln K,$$

fit to all expirations simultaneously subject to no calendar-spread arbitrage between slices, with total variance interpolated across time. The parameters: $a$ = level, $b$ = wing amplitude, $\rho$ = skew ($\in [-1,1]$), $m$ = shift, $\sigma$ = wing curvature. Smoothness and (when $b,\sigma >0,|\rho |<1$) no-strike-arbitrage make it the desk’s favorite pre-processing layer. Gatheral’s book uses it as the input surface from which to study the smile (Ch 3) - the model-independent layer before any SV/LV dynamics are assumed.

The LIBOR Market Model (BM Ch 6). Under its own forward measure each forward-LIBOR is driftless: $d{F}_k={\sigma }_k(t)F_{k}d{Z}_k$, and a caplet is exactly the Black formula (Prop 6.4.1), $\text{Cpl}=P(0,T_i){\tau }_{i}Bl(K,F_i(0),v_i)$. The correlation of the $\{F_k\}$ and the volatility parameterization (piecewise-constant, humped Formulations 6/7, separable) determine everything else.

Cascade calibration (BM Ch 7, Algorithm 7.4.1). With exogenous instantaneous correlations $\rho$ and the approximate swaption-vol formula, the piecewise-constant volatilities ${\sigma }_{k,\beta (t)}$ are recovered one at a time by walking the swaption-volatility matrix (maturity × tenor) left-to-right/top-down; each entry yields a positive algebraic quadratic in exactly one new $\sigma$. Refinements: endogenous interpolation of missing swaption quotes (removes the negative/complex artifacts of naive interpolation), and Monte Carlo reliability tests of the underlying analytical approximation.

Why two-step works (the intuition). Caps fix average vol (a cap is a strip of caplets, each a function of one forward rate’s integrated vol), leaving the differences between rates’ vols - and hence the correlation - to be fixed by swaptions (which depend on a basket of forward rates and their covariance). This decomposition is the rates analogue of “ATM term structure fixes $v_0,\bar{v}$; the skew fixes the rest” in Heston.

---

3. Computational Implementation - a real SVI fit + butterfly check

Fit SVI to a one-maturity smile (SABR-generated, so it has a true smile shape), then check the fitted surface has no strike arbitrage (butterfly $w^{\prime \prime }\ge 0$). Stdlib only.

SVI fits the whole slice to a total-variance RMSE of $2.3\times 10^{-5}$ (about a 0.02-vol-point implied-vol error), and the fitted surface keeps butterfly $w^{\prime \prime }\ge 0$ across the range - no strike arbitrage. This is the layer a desk fits first, then inverts (if it uses LV) or embeds in an SV model (if it uses Heston/SABR): the smooth, arbitrage-free surface is the input to the dynamics, not the dynamics themselves.

---

4. Failure Modes & First-Principles Breakdowns

1. SVI without the arbitrage constraints. A careless SVI fit (free $b,\rho$) can produce $w^{\prime \prime }<0$ locally - strike arbitrage hidden in the “smooth” surface. Always check the butterfly and calendar-spread conditions after fitting, and fit slices jointly to avoid calendar arbitrage.
2. LMM: too many parameters. The full LFM covariance is $O(N^2)$; without low-rank parameterizations (Rebonato’s angles, Schoenmakers–Coffey) or exogenous correlations, the calibration is under-determined. The cascade keeps correlations exogenous and solves the rest one-slice-at-a-time (BM Ch 7).
3. Analytical-approximation risk in rates. The cascade relies on the approximate swaption-vol formula (eq 6.67); BM Ch 8’s Monte Carlo tests show it holds well except in pathological high-vol regimes. Trusting it blindly near the pathology is a calibration error.
4. Conventions compound. Rates add curve/discount conventions (OIS vs LIBOR discounting, day-count, caplet-tenor linkage) on top of the vol conventions of 05; two desks can differ on calibrated ${\sigma }_k$ purely through convention, not model.
5. Surface-first overfitting. Fitting SVI to the entire surface across expirations reduces in-sample error by increasing the number of slices - but the far OTM wings have few quotes, and the fitted wings there are model output, not market fact.

---

5. References

• Gatheral, The Volatility Surface
• Brigo–Mercurio, *Interest Rate Models
• Duffy, Finite Difference Methods in Financial Engineering
• Hull, Options, Futures, and Other Derivatives

---

6. Connected Graph Bridges

• Back: 05 · Failure Modes · Index Hub
• Sibling: Interest Rate & Term Structure Models · Heston & SABR · Implied Volatility Surfaces · Counterparty Risk & XVA
• Base: Stochastic Calculus & Itô