TheQuantBateman

01

INTUITION

Build the state before the equation.

Calibration chooses model parameters that reproduce selected market quotes under an objective function. A low scalar error can hide systematic residuals, unidentifiable parameters and unstable hedges.

01

The objective defines what good fit means.

02

Weights encode liquidity, units and desk priorities.

03

Stability through time is evidence alongside cross-sectional fit.

02

WHY MARKETS CARE

The product exists before the model.

Every surface model and stochastic process requires a repeatable path from cleaned quotes to parameters, diagnostics and fallback decisions.

INSTRUMENTS

vanilla calibration baskets

caps/floors

swaptions

exotic proxy instruments

QUOTE CONVENTION

State calibration instruments, timestamp, bid/offer, objective units, weights, parameter transforms and solver tolerances.

03

MATHEMATICS

Notation, units and exact claims.

θ: parameter vectorq_i: market quotem_i(θ): model quoteW: weight matrixλR(θ): regularizer

QUANT NOTESYNTHETIC · EDUCATIONAL

Weighted least squares

Fit and regularization are explicit modelling choices.

\widehat\theta=\arg\min_\theta\sum_i w_i\left(m_i(\theta)-q_i\right)^2+\lambda R(\theta)

Read the equation together with its financial domain and convention.

QUANT NOTESYNTHETIC · EDUCATIONAL

Gauss–Newton step

The Jacobian reveals local parameter sensitivity and conditioning.

(J^TWJ+\lambda I)\Delta\theta=-J^TWr

Read the equation together with its financial domain and convention.

QUANT NOTESYNTHETIC · EDUCATIONAL

Conditioning

A large condition number signals weak local identification.

\kappa(J^TWJ)=\frac{\lambda_{max}}{\lambda_{min}}

Read the equation together with its financial domain and convention.

04

DERIVATION

Do not jump to the final expression.

DERIVATION

From quote residuals to a parameter update

Linearize model quotes around the current parameter vector.

01
Define residuals
Choose premium, implied-volatility or normalized residuals and preserve their units.
$r_i(\theta)=m_i(\theta)-q_i$
02
Linearize the model
Approximate residuals for a small parameter update using the Jacobian.
$r(\theta+\Delta)\approx r(\theta)+J\Delta$
03
Form the quadratic objective
Insert the linearization into weighted least squares and add damping or regularization.
04
Set the gradient to zero
The normal equations produce a local Gauss–Newton step.
$(J^TWJ+\lambda I)\Delta=-J^TWr$
05
Accept only validated steps
Reprice, enforce domains, inspect residual shape and compare parameters with previous snapshots.
NUMERICAL CHECKConvergence means the optimizer stopped—not that the model is identified.

A calibration result is a parameter vector plus residuals, conditioning, constraints, lineage and a decision on whether it is safe to use.

05

MODEL / PRICING

Fit, compute, then challenge the assumptions.

METHOD

Use bounded parameter transforms, deterministic multi-starts, analytical or validated sensitivities, and a documented fallback.

CALIBRATION

This lesson calibrates calibration itself: choose objective, weights, basket, regularization and stopping rules before fitting the production model.

LIMITATIONS

Non-convex objectives have local minima.

Sparse quotes weakly identify parameters.

Different units and weights produce different optima.

MODEL COMPARISON

Static fit is not dynamics.

Question	Black–Scholes	Local volatility	Heston
Volatility state	One constant σ	σ(S,t) deterministic	vₜ stochastic
Fits today’s surface	No	Exactly, in ideal theory	Approximately by calibration
Forward dynamics	Flat smile	Spot-driven	Variance + correlation driven
Primary strength	Transparent baseline	Vanilla-consistent diffusion	Richer smile dynamics
Primary failure	No smile	Often unrealistic forward skew	Parameter and calibration instability
Compute	Low	Medium: PDE/MC	Medium–high: Fourier/PDE/MC
Hedge implication	Greeks at one σ	State-localized vol hedge	Variance and vol-of-vol risk

Implementation with current QuantLib

Current QuantLib separates market structures, processes, instruments, engines and calibration helpers. Use those abstractions only after the lesson’s conventions, domains and numerical checks are explicit.

API authority: upstream QuantLib reference pinned in the source registry.

06

PYTHON LAB

Theory → implementation → checks.

PYTHON 3 · NUMPY / SCIPY

Linear calibration diagnostic

Solve a weighted linearized fit and inspect conditioning.

REUSABLE EXAMPLE

01from __future__ import annotations

02

03import numpy as np

04

05def gauss_newton_step(jacobian: np.ndarray, residual: np.ndarray, damping: float = 1e-6) -> tuple[np.ndarray, float]:

06 if jacobian.shape[0] != residual.size or damping < 0:

07 raise ValueError("invalid calibration dimensions")

08 normal = jacobian.T @ jacobian + damping * np.eye(jacobian.shape[1])

09 step = np.linalg.solve(normal, -(jacobian.T @ residual))

10 return step, float(np.linalg.cond(normal))

11

12j = np.array([[1.0, 0.1], [1.0, 0.4], [1.0, 0.9]])

13r = np.array([0.01, -0.005, 0.002])

14step, condition = gauss_newton_step(j, r)

15assert np.isfinite(step).all() and np.isfinite(condition)

16assert np.linalg.norm(j @ step + r) < np.linalg.norm(r)

17print(np.round(step, 6), round(condition, 2))

07

INTERACTIVE LAB

Move the state. Challenge the equation.

OPTIMIZER PATH AND RESIDUAL GEOMETRY

Calibration residual lab

Change weights, noise and initial parameters; watch optimizer path, residual heatmap and conditioning.

SYNTHETIC · CONTROLLED SCENARIOS

SCENARIO INTENSITY70%

Final RMSE0.0016

Condition stateWell identified

Iterations30

08

FRONT OFFICE

Where the model meets the book.

ON THE DESK

“A perfect fit with unstable parameters is an expensive interpolation scheme.”

VISIBLE INPUTS

timestamped quotes

bid/offer

basket

weights

constraints

CALIBRATION

This lesson calibrates calibration itself: choose objective, weights, basket, regularization and stopping rules before fitting the production model.

RISK

parameter jump

basis residual

solver risk

model selection

DAILY WORKFLOW

freeze snapshot
clean basket
multi-start fit
inspect residuals
approve/fallback

Production failure modes

unit mismatch
silent constraint binding
nondeterministic start
missing diagnostics

09

MACRO CONNECTION

Map the transmission channel.

MACRO CONNECTION

Regime shifts expose parameter identification

When the surface changes shape, parameters that were redundant in calm markets can move abruptly or hit bounds.

Regime shift

changes quote geometry

Calibration basket

exposes sensitivities

Parameters

move or become unstable

Model governance

accepts, regularizes or falls back

10

COMMON PITFALLS

Most failures begin outside the formula.

01

Reporting only RMSE.

02

Mixing premium and vol residuals without normalization.

03

Using one starting point.

04

Ignoring day-over-day parameter stability.

11

SOURCES / FURTHER READING

Attribution with implementation authority.

researchBSD-3-Clause

Volatility, Monte Carlo and stochastic-volatility lectures

Research map for the mathematical progression and numerical experiments; prose, examples and code are original.

Source: Computational Finance Course
Author: L. A. Grzelak
Ref: main

OPEN ORIGINAL SOURCE ↗LICENSE ↗

implementation referenceQuantLib permissive license

Current volatility structures, processes, calibration helpers and tests

Implementation reference for production abstractions and validation patterns.

Source: QuantLib upstream
Author: QuantLib contributors
Ref: v1.42.1

OPEN ORIGINAL SOURCE ↗LICENSE ↗

Weighted least squares

Gauss–Newton step

Conditioning

From quote residuals to a parameter update

Define residuals

Linearize the model

Form the quadratic objective

Set the gradient to zero

Accept only validated steps

Volatility, Monte Carlo and stochastic-volatility lectures

Current volatility structures, processes, calibration helpers and tests