Pricing Models — Cross-Cutting Comparison

This note compares every derivatives-pricing model and numerical method referenced across the Finance library — Black-Scholes-Merton, local-vol (Dupire), stochastic-vol (Heston/SABR/SVI), jump-diffusion (Merton/Kou/Bates), Lévy processes (CGMY/VG/NIG), local-stochastic-vol (LSV), PDE solvers, Monte Carlo, binomial/trinomial trees, least-squares MC (Longstaff-Schwartz), Fourier methods (Carr-Madan), COS method (Fang-Oosterlee), quasi-Monte Carlo (Sobol/Halton), finite-difference, and deep-learning vol surfaces (Horvath-Muguruza-Tomas). Each row maps to where it sits on model complexity, calibration cost, smile fit, term-structure fit, Greek accuracy, and computational cost. Read the tables; decision tree at the end.

See also

1. The model taxonomy

ANALYTIC / SEMI-ANALYTIC                    SIMULATION / NUMERICAL
  Black-Scholes-Merton (1973)                 Monte Carlo (Boyle 1977)
  Bachelier (1900, "normal model")            Quasi-Monte Carlo (Sobol, Halton, Niederreiter)
  Hull-White short-rate                       Longstaff-Schwartz LSM (2001)
  Vasicek                                     Stochastic mesh (Broadie-Glasserman 1997)
  CIR                                         Particle methods (Guyon-Henry-Labordère)
  HJM (special cases)                         Tree (binomial Cox-Ross-Rubinstein 1979,
  Black-Karasinski                             trinomial Boyle 1986)
                                               Finite-difference (Crank-Nicolson, ADI)
                                               Fourier (Carr-Madan 1999)
                                               COS method (Fang-Oosterlee 2008)
                                               Heath-Jarrow-Morton MC

LOCAL VOL / STOCHASTIC VOL / JUMP            HYBRID / RATES + EQUITY
  Dupire local vol (1994)                      LSV (local-stochastic vol)
  Heston SV (1993)                             Heston + Hull-White
  SABR (Hagan et al 2002, 2014 fix)            LMM/BGM (LIBOR Market Model)
  SVI (Gatheral 2004; eSSVI Gatheral-Jacquier 2014) Cheyette (Markov-functional HJM)
  Merton jump-diffusion (1976)                 SABR-LMM
  Kou double-exponential JD (2002)
  Bates SV-jump (1996)                       MACHINE-LEARNING
  CGMY Lévy (Carr-Geman-Madan-Yor 2002)        Horvath-Muguruza-Tomas DNN (2021)
  Variance Gamma (Madan-Seneta 1990)           Neural SDE
  Normal Inverse Gaussian (Barndorff-Nielsen 1997) Differentiable solvers
  Mixed local-stochastic vol (LSV)             Deep hedging (Buehler et al 2019)
                                               Sig-payoffs (Lyons rough-path)

2. The models compared on complexity, calibration, smile fit

ModelYearComplexity (state dim)Calibration costSmile fitTerm-structure fitGreeks
Bachelier19001 (normal price)trivialnone (flat smile = absent)staticanalytic
Black-Scholes-Merton19731 (S only)trivialnone (flat)staticanalytic
Dupire local vol19941 (S, σ_loc(S, t))exact reproduction of smile (PDE inversion of Dupire’s equation)exact (perfect by construction)exact static, poor dynamics (sticky-strike implied)numerical PDE
Heston SV19932 (S, v)semi-closed via Carr-Madan/Liptongood (curved smile)staticsemi-analytic
SABRHagan 20022 (F, σ)Hagan asymptotic formula (smile arc)excellent for short-dated FX/ratessmile-only, term separateasymptotic
SVI / SSVI / eSSVIGatheral 2004+parametric (5 params/slice)nonlinear least squaresexcellentpiecewise (one slice per expiry)numerical
Merton jump-diffusion19761 + Poissonsemi-closed (series sum)good for short-dated (jump risk premium)staticanalytic series
Kou JD20021 + double-exp jumpssemi-closedgoodstaticsemi-analytic
Bates SV-jump19963 (S, v, J)Carr-Madanvery goodstaticsemi-analytic
CGMY2002Lévy, 4 paramsCarr-Madanexcellent (controls skew + kurtosis)piecewise via term-structuresemi-analytic
Variance Gamma1990Lévy, 3 paramsCarr-Madangoodpiecewisesemi-analytic
NIG1997Lévy, 4 paramsCarr-Madangoodpiecewisesemi-analytic
LSV (Heston + local vol leverage)2010+2 + leverage functioniterative — Heston SV + Dupire layerexcellentgood (mixed dynamics)numerical
Hull-White19901 (r)bond curve + cap/swaption surfacen/ayes, by constructionsemi-analytic
LMM / BGM1996/1997n (one rate per tenor)swaption surface + cap surfaceyes (rate smile)yesMC primarily
Cheyette (Markov-functional HJM)1992low-dim Markov stateswaption surfaceyesyessemi-analytic
Horvath-Muguruza-Tomas DNN (2021)2021parametric (DNN weights)DNN fitted once to model outputexcellent (matches Heston/SABR/etc.)depends on training setautograd
Neural SDE2020+learned drift + diffusionadjoint sensitivitytrained to datatrainedautograd

3. Numerical methods — when to use which

MethodProsConsWhen to use
Closed-form (BSM, Bachelier, Black ‘76)instant, exact, analytic Greeksrestrictive assumptionsvanilla European on simple dynamics
Binomial tree (Cox-Ross-Rubinstein 1979)intuitive, handles American, early-exerciseslow convergence (1/N), only one risk factor easilyteaching, simple American
Trinomial tree (Boyle 1986)better convergence than binomialstill 1DAmerican w/ better Greeks
Implicit / Crank-Nicolson finite-differencestable, accurate for low-dim PDEcurse of dimensionality (≤ 3D practical)Heston/local-vol PDE, exotic w/ barriers
ADI / Strang-splittingmulti-dim FDrequires sparse linear algebraLSV, Heston-HW hybrid
Monte Carloany dim, any payoff, path-dependent nativeslow convergence (1/√N)exotics, path-dependent, high-dim
Quasi-Monte Carlo (Sobol, Halton)1/N convergence rate (vs MC’s 1/√N) for smooth integrandsbreaks for non-smooth (some barrier options)high-dim smooth
Multi-level MC (Giles 2008)variance reduction via telescopingextra implementationSDEs with strong-convergence
Longstaff-Schwartz LSM (2001)American on MC (polynomial regression on continuation value)basis function choice mattersBermudan / American on complex dynamics
Stochastic mesh (Broadie-Glasserman 1997)exact bounds for Americanexpensivehigh-fidelity American
Fourier transform (Carr-Madan 1999)exploit characteristic function (Lévy, affine)requires affine modelHeston, Bates, Lévy, CGMY, VG
COS method (Fang-Oosterlee 2008)faster than Carr-Madan, cleaner truncationrequires bounded characteristicsame as Fourier
Particle methods (Guyon-Henry-Labordère)LSV with exact LV matchingimplementation-heavyLSV in production
Deep PDE solvers / PINNshigh-dim PDErequires trainingresearch
Differentiable MC (autograd)Greeks for freerequires PyTorch/JAX rewritegrowing
GPU MC100× speedup for large product booksCUDA / Triton skillproduction dealer books

4. The Fourier / COS family — the workhorse for affine + Lévy

For models with a known characteristic function (BSM, Heston, Bates, Merton JD, CGMY, VG, NIG):

  • Carr-Madan 1999 — option price via Fourier transform of damped call: with the characteristic function of . FFT-able; one model evaluation gives the entire strike range.
  • COS method (Fang-Oosterlee 2008) — Fourier-cosine series on a truncated domain. Same idea, better convergence and easier error control. The de-facto standard 2010s+ for Lévy and Heston vanilla pricing.
  • Lewis 2001 — slightly different transform; cleaner for some structured products.

For non-vanilla payoffs (barriers, Asians) one combines these with Monte Carlo or path-Fourier methods.

5. The exotic ladder — what works for what

PayoffBest model + numerical method
Vanilla EuropeanBSM + closed-form OR Heston + Carr-Madan
Vanilla American (equity)Heston + LSM or Crank-Nicolson PDE
Vanilla American (rates)LMM + LSM or Markov-functional
Barrier (knock-in/out)local vol + PDE; LSV + MC; closed-form BSM for simple
Asian (arithmetic)MC with control variate; Curran approximation for closed-form quasi-analytic
Lookback (max/min)PDE in 2D (S, M) or MC; Goldman-Sosin-Gatto formula for closed-form BSM
Cliquet / ratchetMC with LSV; vega-sensitive
Forward-startHeston + Carr-Madan; vol-of-vol matters
Volatility swap, variance swapstatic replication via vanilla strip (Carr-Madan 2001)
VIX future, VIX optionHeston / Bates + MC; or fit VIX vol surface directly
AutocallableLSV + MC; LSM for early-redemption; key for retail structured note (Asia, Europe)
Accumulator (“I-kill-you-later”)LSV + MC; massive vega-of-vega risk
Reverse convertibleBSM or Heston + closed-form (short put-like)
Principal-protected noterates model + closed-form on call piece
CMS spreadLMM w/ smile + MC; vega-of-vega + correlation; Brigo-Mercurio
Bermudan swaptionLMM + LSM or Markov-functional + PDE
Range accrualHull-White or LMM + MC
TARN (Target Redemption Note)LMM + MC; early-termination tail risk
Catastrophe bondPoisson + lognormal severity + MC
Mortgage prepaymentstructural prepayment model + interest-rate MC; OAS engine
CDO / CLOfactor copula (Gaussian, Student-t, Marshall-Olkin) + MC; base-correlation surface
CDS index tranchebase-correlation surface + Gaussian copula + MC
Compound option (call on call)Geske 1979 closed-form BSM; otherwise MC
Quantodual-currency + correlation; quanto adjustment via Girsanov
Multi-asset basketGaussian or Student-t copula + MC; PDE for low-dim
Best-of / worst-ofLSV + MC; correlation skew matters

6. Calibration — the fitting problem

ModelWhat to fit toFitting method
BSMa single ATM vol pointtrivial
Local volthe entire implied-vol surfaceDupire formula gives σ_loc(K, T) by inversion
Hestonsmile + term structure (5 params: v0, θ, κ, ξ, ρ)nonlinear least squares on vanilla prices via Carr-Madan
SABRper-slice smile (4 params: α, β, ρ, ν)Hagan formula + nonlinear LS
SVI / eSSVIper-slice smile (5 params SVI / 2-3 SSVI)nonlinear LS; arbitrage-free constraints
Batessmile + skew (Heston + jump)nonlinear LS via Carr-Madan
CGMY / VG / NIGsmile + termCarr-Madan inversion
Hull-Whitebond curve + cap/swaption ATMcalibration to swaption surface
LMM/BGMswaption surface + cap surfaceglobal LS on calibrated swaption + cap
LSVlocal-vol leverage to vanilla + SV dynamics to exoticiterative (Heston seed → LV adjustment)
Horvath DNNoffline-trained on model outputone-time training; inference is fast

The 2024 frontier: Joint calibration of vol + rates (Brigo-Mercurio book is the bible), machine-learning-accelerated calibration (Horvath et al 2021 Quant Finance; Liu-Oosterlee-Bohte 2019), differentiable calibration via PyTorch/JAX gradients.

7. Smile-fit quality on the SPX vol surface

Empirical observation (Gatheral Volatility Surface 2nd ed; Sepp & Yashin 2022 Quant Finance):

ModelFront-month smile (1W–1M)Mid (3M–6M)Long-dated (1Y+)Behavior in extreme strikes
BSMflat (none)flat (none)flat (none)n/a
Local volexactexactexactsticky-strike (unrealistic for surface dynamics)
HestonOK (smile too symmetric)goodgoodpoor in wings
SABRexcellentgoodneeds hump correctionOK
Batesexcellentexcellentgoodgood
LSV (Heston + local)excellentexcellentexcellentgood
CGMY / VGexcellentgoodgoodexcellent in wings

Local vol is the right answer for today’s smile; LSV / SV is the right answer for how the smile evolves. Vanna-volga (Castagna-Mercurio 2007) is a pragmatic interpolation for FX desks.

8. Computational cost — order-of-magnitude

OperationBSMHeston (Carr-Madan)SABR (asymptotic)LSV PDELSV MCLSV deep-hedging
One vanilla priceµsmsµs10 ms100 msµs (after training)
One Greek (delta)µsmsµs (perturb)10 ms100 ms (pathwise)µs
One vegaµsmsµs10 ms (perturb)100 ms (likelihood ratio)µs
One barrier optionms (analytic)100 msn/a100 ms1 sµs (trained)
One Bermudan (LSM)n/an/an/a100 ms PDE10 s LSMn/a
One TARNn/an/an/an/a1 minn/a (specialized)
One CDO tranchen/an/an/an/a1 sn/a
Daily revaluation of dealer book (10⁵ trades)n/an/an/aminuteshoursminutes

A modern dealer’s overnight grid (revalue tens of thousands of trades) typically uses GPU Monte Carlo for exotics, PDE for European/American on small dynamics, and Carr-Madan / COS for vanillas in batch.

9. Fixed income — separate world

Equity-options models (Heston, Dupire, etc.) do not apply directly. Rates products are priced via:

ModelWhenNotes
Vasicektoy / teachingmean-reverting; can hit negative rates
CIRtoy / teachingnon-negative; affine
Hull-White (extended Vasicek)swap/cap/swaption with rate curveanalytic for bond, cap; semi for swaption
Black-Karasinskicap, swaptionnon-negative
HJM (in general)researchcurve modeled directly
LIBOR Market Model (BGM 1996/1997)cap + swaption surface; modern industry standardcalibrate to swaption surface; MC pricing
SABR-LMMsmile-extended LMMrates smile + LMM dynamics
Cheyette / Markov-functional HJMlow-dim Markov statesemi-analytic, dealer-friendly
Two-factor short-rate (G2++)hybridflexible
Multi-curve framework (post-2008 OIS / IBOR split)mandatory for valuing collateralizeddiscount on OIS, project IBOR forwards
SOFR transition models (post-2021 LIBOR cessation)post-LIBORterm SOFR (CME), backward-looking compound SOFR

See fixed-income-deep for the full rates stack.

10. Credit pricing

ModelWhen
Merton structural (1974)corporate credit risk; equity = call on firm value
Black-Cox (1976)barrier extension
Reduced-form intensity (Jarrow-Turnbull 1995, Lando 1998, Duffie-Singleton 1999)CDS pricing; default = Poisson w/ intensity
CreditMetricsportfolio loss; transition matrix-based
CreditRisk+ (CSFB)actuarial Poisson
Gaussian copula (Li 2000)CDO pricing — infamous for 2008
Marshall-Olkin / random factor loadingpost-2008 alternative copulas
Base-correlation surfacedealer-grade CDO tranche
Random-recovery model (Andersen-Sidenius-Basu 2003)CDO w/ recovery uncertainty
Hawkes processesclustered defaults

11. Machine-learning pricing models — the 2020–2026 wave

MethodWhen
Horvath-Muguruza-Tomas 2021 “Deep learning vol surfaces”offline-train DNN to take (model params) → (vol surface); 1000× faster inference for calibration
Liu-Oosterlee-Bohte 2019DNN as a fast surrogate for Heston Carr-Madan
Buehler-Gonon-Teichmann-Wood 2019 “Deep hedging”learn hedging strategy directly; no PDE; handles transaction cost natively
Becker-Cheridito-Jentzen-Welti 2019DNN for American option pricing
Sirignano-Spiliopoulos 2018 “Deep Galerkin”DNN solves high-dim PDE
Han-Jentzen-E 2018 “Deep BSDE”backward stochastic differential equation solver
Lyons rough-path / signature payoffs (Lyons, Salvi-Pannier-Lyons 2021)path-dependent payoffs as functions of signatures
Neural SDE (Kidger-Foster-Li-Lyons 2021)learn drift + diffusion from price data
Differentiable Monte Carloautograd through MC for fast Greeks
Differentiable PDE solvers (JAX-FD, jax-pde)gradient w.r.t. boundary conditions, params

12. The post-2008 valuation-adjustment (xVA) stack

For OTC derivatives, the all-in price includes:

AdjustmentWhat it covers
CVA (Credit Valuation Adjustment)counterparty default risk
DVA (Debit VA)your own default risk (controversial; subtract from CVA)
FVA (Funding VA)funding cost of uncollateralized position
ColVA (Collateral VA)imperfect collateralization
KVA (Capital VA)regulatory capital required for the trade
MVA (Margin VA)initial margin funding cost
AVA (Additional VA)prudent valuation adjustment under EBA

Each is a high-dim expectation over rates + credit + equity scenarios, typically computed by GPU MC on a global netting set. The dealer-grade xVA engine (Murex, Numerix, Quantifi, Calypso) costs millions in licensing.

13. Decision tree — pick by product + accuracy + budget

What's the product?
├─ Vanilla European, liquid
│    → BSM closed-form (1 µs) or Heston + Carr-Madan if smile matters
├─ Vanilla American, equity
│    ├─ Quick → Cox-Ross-Rubinstein binomial tree
│    ├─ Better → Crank-Nicolson PDE
│    └─ Complex dynamics → Longstaff-Schwartz LSM
├─ Barrier (knock-out / knock-in)
│    ├─ Simple → BSM closed-form (Rich)
│    ├─ Smile → local vol + PDE
│    └─ Smile + dynamics → LSV + MC
├─ Asian (path-dependent average)
│    ├─ Geometric → closed-form
│    └─ Arithmetic → MC + control variate (Kemna-Vorst geometric)
├─ Bermudan / autocallable / TARN
│    └─ LSV + MC + Longstaff-Schwartz LSM
├─ VIX option / variance swap
│    ├─ Variance swap → static replication via vanilla strip (Carr-Madan 2001)
│    └─ VIX option → Heston/Bates + MC
├─ FX option (short-dated)
│    └─ SABR + Hagan asymptotic
├─ FX exotic (long-dated, smile-sensitive)
│    └─ LSV + MC or Vanna-Volga
├─ Rates cap, floor, swaption
│    ├─ ATM only → Black '76 + Hull-White
│    └─ Smile → SABR-LMM + MC
├─ CMS spread, range accrual, callable bond
│    └─ LMM + MC + LSM for early exercise
├─ Mortgage-backed security
│    └─ structural prepayment + Hull-White rate model + MC; OAS engine
├─ CDO / CLO tranche
│    └─ base-correlation surface + Gaussian copula + MC
├─ Catastrophe bond
│    └─ Poisson frequency + lognormal severity + MC
├─ ADC / weather derivative
│    └─ custom Poisson / mean-reverting OU + MC
├─ Crypto option (high vol, jumps)
│    └─ Bates SV-jump or Lévy + Carr-Madan
└─ Production xVA engine (dealer book)
     └─ GPU MC on Heston/Hull-White hybrid; xVA accounting layer

14. Anti-patterns

  1. BSM for short-dated equity options near earnings — vol smile + jumps mandate at least Heston, often Bates.
  2. Local vol for forward-vol-sensitive products — sticky-strike dynamics give wrong forward vol; use SV or LSV.
  3. Constant-vol Heston for long-dated — calibrate term structure of v0, θ; or use term-structured Heston.
  4. MC without variance reduction for vanilla — use control variate or QMC.
  5. PDE for high-dim — > 3 dim is impractical; use MC.
  6. Pricing a Bermudan with European MC — use LSM or PDE.
  7. Ignoring xVA for OTC trades — material in modern post-2010 P&L.
  8. Calibrating to OTM with poor data — wings are noisy; use SSVI / eSSVI for arbitrage-free interpolation.
  9. Calibrating to today’s smile only — without time-series fit, dynamics will be wrong.
  10. Using a Gaussian copula for tail-risk credit — 2008 lesson; use t-copula or Marshall-Olkin.

15. The 2024–2026 frontier

  • GPU + JAX/PyTorch differentiable pricing — Cantor, Goldman, JP Morgan, BlackRock all moving toward differentiable MC for daily P&L attribution.
  • Rough-vol models (Gatheral-Jaisson-Rosenbaum 2018 “Volatility is rough”) — vol is best modeled as fractional Brownian motion with H ≈ 0.1; rough Bergomi, rough Heston.
  • rough Heston / rough Bergomi — improved smile fit short-dated, especially equity.
  • Sig-payoffs — Lyons rough path signatures as universal approximators of path-dependent payoffs.
  • Deep hedging (Buehler et al 2019) at Bank of America, JP Morgan — learn hedging strategy under realistic constraints (frictions, capital).
  • PINN / neural-PDE for high-dim — Sirignano-Spiliopoulos, Han-E.
  • xVA at the cell level (Murex, Numerix, Adaptiv) — netting-set level XVA, post-trade analytics in seconds.
  • SOFR transition models — post-LIBOR rates with overnight compounding base curve.
  • Crypto derivatives modeling — Bates / Lévy plus weekend-effect dummies; high-frequency funding-rate models.
  • AI-aided model validation — replicate independent pricing by transformer-based DNN; back-test for arbitrage.

Adjacent

When to pick what

The fastest narrowing: liquid vanilla → BSM; smile matters → Heston + Carr-Madan; smile + path dependence → LSV + MC; Bermudan → LSM; barrier → local vol + PDE; rates → Hull-White or LMM; CDO → base-correlation + Gaussian copula (with t-copula for tails); xVA → GPU MC. The single biggest practical lesson of the last 25 years (post-1998 LTCM, post-2008 GFC, post-2020 COVID) is calibrate to today’s surface but stress against alternative dynamics — local vol is correct for today, wrong for tomorrow, and a model that fits perfectly today is not a model of forward dynamics. Pick the simplest model that captures the relevant risk; price under the chosen model; stress-test under alternatives.