Monte Carlo for risk estimation: from sampling to quantum

Created on May 15, 2026

2026   ·   math   quantum-computing   risk   finance   vertical   ·   math   probability-finance   probability   finance

\[\newcommand{\E}{\mathbb{E}} \newcommand{\Risk}{\mathrm{Risk}}\]

The thesis, in risk terms

A risk model is only ever as good as the loss it can quantify. Strip away the dynamics and the story and the question that remains is the same one: what is the expected — and tail — loss this system produces? Stated that plainly, risk modeling has a single structural problem. The loss functional you actually care about is almost never available in closed form, and the part you care about most — the tail, the rare adverse alignment, the flash event — is exactly the part where closed-form approximations are least trustworthy.

Monte Carlo is the response to that problem, and the history of risk modeling is largely the history of taking it seriously. This survey runs that history through a risk-analysis lens — loss exposure, not algorithms — and then asks where two more recent answers fit: amortized variational methods (VAE) and quantum amplitude estimation. Field first; my own work makes no appearance.

What a risk model must deliver

Before simulating anything, the object being estimated has to be a coherent notion of risk. The axioms of (Artzner et al., 1999) fix what a risk measure must satisfy (monotonicity, sub-additivity, positive homogeneity, translation invariance) and, in doing so, expose why variance and plain value-at-risk mislead precisely in the tail. The operational consequence is that a risk specification has to connect adaptation and decisions to realized loss exposure — not to a convenient surrogate.

Once the tail is the quantity of interest, the relevant functional is a conditional-tail object. (Rockafellar & Uryasev, 2000) made Conditional Value-at-Risk tractable as an optimization target: instead of treating CVaR as a post-hoc diagnostic, they rewrote it as a convex problem you can actually solve. That move is the unsung reason finance, insurance, and a growing slice of ML uses CVaR rather than VaR as the default tail metric — VaR tells you a threshold, CVaR tells you the conditional loss past it, and only one of those survives optimization.

Why the loss has to be simulated

If the loss integral were closeable, none of the simulation machinery would exist. It generally is not, so the field estimates it. The estimator itself originates with (Metropolis & Ulam, 1949): replace an intractable expectation with a sample average and accept estimator variance as the price of admission. The Markov-chain construction that made this work in high-dimensional, dependent state is (Metropolis et al., 1953), generalized into the sampling framework still used today by (Hastings, 1970). Any time you hear “Markov risk model,” that lineage is what the word Markov is doing.

The modern, risk-facing consolidation is textbook by now — (Rubinstein & Kroese, 2016) for the estimator and its variance, (Glasserman, 2004) for the financial-risk specialization where the tail is the deliverable. The recurring lesson across both: a naive Monte Carlo estimate of a tail loss is dominated by the region it almost never samples, so variance reduction and importance sampling are not optimizations — they are the difference between an estimate and noise.

Monte Carlo vs Variational Autoencoders — estimating vs amortizing

For a long time the only honest answer to “the integral has no closed form” was Monte Carlo. The variational autoencoder line of (Kingma & Welling, 2014) and (Rezende et al., 2014) added a second answer that the risk community has been slow to take seriously: replace the intractable posterior with a learned approximation, then sample from that. The trade is precise and worth stating in risk terms.

  • Monte Carlo estimates an expectation directly. It is unbiased; its variance is what you pay; the standard error is \(\sigma/\sqrt{N}\) in \(N\) samples. For a rare-event tail, that variance is brutal — most of your samples carry no information about the quantity you care about, which is why (Glasserman, 2004) spends most of its pages on importance sampling and control variates.
  • VAE estimates a distribution parameterized by a neural encoder–decoder. It is biased by the choice of variational family; in exchange, the variance of any downstream expectation drops sharply because the approximation already concentrates mass where it matters. Once the encoder is trained, sampling is \(O(1)\) — the cost was amortized.

The honest decomposition is therefore not “MC vs VAE” but “where is your budget?” If the integral is one-shot and you cannot tolerate bias, MC with variance reduction is still correct. If you will evaluate similar expectations many times — pricing across strikes, risk across portfolios, posterior queries across scenarios — the amortization VAE buys eats the MC sample cost alive. The trap is symmetric: an MC user who pretends naive estimators handle the tail, and a VAE user who pretends the variational gap is zero. Both are common; neither survives a serious tail diagnostic.

Quantum amplitude estimation — the quadratic speedup that matters

The third answer is the one the field is still partly arguing about. The core observation is older than the hype: (Brassard et al., 2002) showed that amplitude estimation — a quantum subroutine that estimates the probability of a marked outcome — converges with error \(O(1/N)\) in \(N\) queries, where classical Monte Carlo converges as \(O(1/\sqrt{N})\). That is a quadratic speedup, and it is not asymptotic hand-waving: it applies to the same expectation a financial-risk MC actually computes.

(Montanaro, 2015) carried this through to generic Monte Carlo: any expectation accessible to classical MC admits the quadratic speedup under a quantum oracle for the underlying distribution. The applied line landed shortly after: (Rebentrost et al., 2018) laid out the specifically-financial case (option pricing, credit risk), and (Stamatopoulos et al., 2020) demonstrated option pricing on actual quantum hardware with the amplitude-estimation primitive.

Read as risk analysis, the implication is narrow but real. A quadratic-in-precision speedup means estimating a tail quantile to one extra digit costs the square root of the classical sample count — not a constant factor saving but a different scaling regime. The honest caveat is equally narrow: you need a quantum oracle for the loss distribution, and loading classical data into amplitude form is its own hard problem. The literature now spends as much effort on data loading as on the amplitude-estimation primitive itself, because that loading cost is what decides whether the quadratic speedup is realized in practice. The field hasn’t settled it. But “MC, but quadratically tighter in precision” is now a real third option, not vaporware.

Key references — 5-field cards

Each card: Claim · Method · Matters (the risk-analysis reason it is here) · Connects to (where it sits in the survey) · Code.

[1] (Metropolis & Ulam, 1949) — The Monte Carlo Method

  • Claim: an intractable expectation can be replaced by a sample average.
  • Method: stochastic sampling of the loss functional.
  • Matters: the entire reason a non-closeable risk integral is estimable.
  • Connects to: every section below — the root of the family.
  • Code: — (1949 historical paper; modern MC libraries are reimplementations).

[2] (Metropolis et al., 1953) — Equation of State by Fast Computing Machines

  • Claim: dependent high-dimensional state can be sampled via a Markov chain.
  • Method: the Metropolis acceptance rule.
  • Matters: makes simulation feasible when state is correlated over time.
  • Connects to: any “Markov risk model” — this is the origin of the word.
  • Code: — (historical; PyMC / Stan / emcee reimplement it).

[3] (Hastings, 1970) — MC Sampling Using Markov Chains

  • Claim: generalizes Metropolis to arbitrary target distributions.
  • Method: Metropolis–Hastings.
  • Matters: the formal basis of Markov-chain risk simulation.
  • Connects to: state-dependent risk processes; ergodic-average estimators.
  • Code: — (historical; standard in every MCMC library).

[4] (Artzner et al., 1999) — Coherent Measures of Risk

  • Claim: axioms a sound risk measure must satisfy.
  • Method: the coherence axiom set (mono / sub-add / hom / transl-inv).
  • Matters: explains why variance and VaR mislead in the tail.
  • Connects to: motivates a tail-aware reading of any \(\Risk(\cdot)\).
  • Code: — (theory paper; trivial to encode the axioms as predicates).

[5] (Rockafellar & Uryasev, 2000) — Optimization of CVaR

  • Claim: conditional tail loss is a tractable optimization target.
  • Method: convex reformulation of CVaR.
  • Matters: gives the flash-risk tail a workable definition.
  • Connects to: every objective that targets a conditional tail.
  • Code: — (standard convex-solver implementations exist; e.g. cvxpy).

[6] (Glasserman, 2004) — Monte Carlo Methods in Financial Engineering

  • Claim: tail risk is estimable but only with variance control.
  • Method: importance sampling, control variates, stratification.
  • Matters: the practical core of estimating a flash-risk tail.
  • Connects to: the rare-event section; the whole reason naive MC fails.
  • Code: — (book; reference implementations widely available).

[7] (Rubinstein & Kroese, 2016) — Simulation and the Monte Carlo Method

  • Claim: the MC estimator and its variance, end to end.
  • Method: estimator construction + rare-event simulation.
  • Matters: why naive tail simulation is noise, not an estimate.
  • Connects to: the \(\sigma/\sqrt{N}\) argument behind variance reduction.
  • Code: — (textbook; companion code by edition).

[8] (Kingma & Welling, 2014) — Auto-Encoding Variational Bayes

  • Claim: a learned posterior approximation can replace MCMC for many inference tasks.
  • Method: reparameterization trick + amortized encoder.
  • Matters: turns repeated MC integrals into \(O(1)\) sampling after training.
  • Connects to: the MC-vs-VAE trade; biased-but-amortized vs unbiased-but-expensive.
  • Code: ✓ — original implementation and many open-source clones (PyTorch / JAX).

[9] (Rezende et al., 2014) — Stochastic Backpropagation and Deep Generative Models

  • Claim: gradients of expectations under deep generative models are tractable.
  • Method: stochastic backprop through the sampling layer.
  • Matters: the reason VAE-style estimators are end-to-end optimizable.
  • Connects to: the variance side of the MC-vs-VAE trade.
  • Code: ✓ — reference implementations in every deep-generative-model library.

[10] (Brassard et al., 2002) — Quantum Amplitude Amplification & Estimation

  • Claim: quantum amplitude estimation gives \(O(1/N)\) error in \(N\) queries.
  • Method: Grover-style amplification + phase-estimation readout.
  • Matters: the quadratic-in-precision speedup over classical MC.
  • Connects to: every “quantum MC” paper that follows.
  • Code: — (theory; primitives in Qiskit / Cirq).

[11] (Montanaro, 2015) — Quantum Speedup of Monte Carlo Methods

  • Claim: any classical MC expectation inherits the quadratic speedup.
  • Method: reduces generic MC to amplitude estimation.
  • Matters: makes the speedup a general MC theorem, not a special case.
  • Connects to: the quantum section’s load-bearing claim.
  • Code: — (theory; reference circuits in Qiskit Finance).

[12] (Rebentrost et al., 2018) — Quantum Computational Finance: MC Pricing of Derivatives

  • Claim: option pricing inherits the amplitude-estimation speedup.
  • Method: encodes payoff + path simulation into a quantum oracle.
  • Matters: the financial-risk-specific instantiation of Montanaro.
  • Connects to: any actual finance-side MC use case.
  • Code: ✓ — reference implementations in Qiskit Finance.

[13] (Stamatopoulos et al., 2020) — Option Pricing on Quantum Computers

  • Claim: amplitude-estimation option pricing runs on actual hardware.
  • Method: experiments on real superconducting devices.
  • Matters: pulls the speedup story out of theory and into measurement.
  • Connects to: the honest “data loading is the bottleneck” caveat.
  • Code: ✓ — code released alongside the Qiskit Finance module.

References are real, established works (verify exact fields before citing formally). No DOIs are asserted. Code lines marked ✓ where I am confident a reference implementation is publicly available; “—” where the work is foundational/historical and the modern code is reimplementation, not the authors’ release.

References

2020

  1. Option Pricing using Quantum Computers
    Nikitas Stamatopoulos, Daniel J. Egger, Yue Sun, and 4 more authors
    Quantum, 2020
    Verify fields before relying on them
    arXiv

2018

  1. Quantum Computational Finance: Monte Carlo Pricing of Financial Derivatives
    Patrick Rebentrost, Brajesh Gupt, and Thomas R. Bromley
    Physical Review A, 2018
    Verify fields before relying on them
    arXiv

2016

  1. Simulation and the Monte Carlo Method
    Reuven Y. Rubinstein and Dirk P. Kroese
    2016
    Verify edition/year before relying on them

2015

  1. Quantum Speedup of Monte Carlo Methods
    Ashley Montanaro
    Proceedings of the Royal Society A, 2015
    Verify fields before relying on them
    arXiv

2014

  1. Auto-Encoding Variational Bayes
    Diederik P. Kingma and Max Welling
    In International Conference on Learning Representations (ICLR), 2014
    arXiv 2013; ICLR 2014. Verify before relying on them
    arXiv
  2. Stochastic Backpropagation and Approximate Inference in Deep Generative Models
    Danilo Jimenez Rezende, Shakir Mohamed, and Daan Wierstra
    In International Conference on Machine Learning (ICML), 2014
    Verify fields before relying on them
    arXiv

2004

  1. Monte Carlo Methods in Financial Engineering
    Paul Glasserman
    2004
    Verify fields before relying on them

2002

  1. Quantum Amplitude Amplification and Estimation
    Gilles Brassard, Peter Høyer, Michele Mosca, and 1 more author
    In Contemporary Mathematics: Quantum Computation and Information, 2002
    Verify fields before relying on them
    arXiv

2000

  1. Optimization of Conditional Value-at-Risk
    R. Tyrrell Rockafellar and Stanislav Uryasev
    Journal of Risk, 2000
    Verify fields before relying on them

1999

  1. Coherent Measures of Risk
    Philippe Artzner, Freddy Delbaen, Jean-Marc Eber, and 1 more author
    Mathematical Finance, 1999
    Verify fields before relying on them

1970

  1. Monte Carlo Sampling Methods Using Markov Chains and Their Applications
    W. K. Hastings
    Biometrika, 1970
    Verify fields before relying on them

1953

  1. Equation of State Calculations by Fast Computing Machines
    Nicholas Metropolis, Arianna W. Rosenbluth, Marshall N. Rosenbluth, and 2 more authors
    The Journal of Chemical Physics, 1953
    Verify fields before relying on them

1949

  1. The Monte Carlo Method
    Nicholas Metropolis and Stanislaw Ulam
    Journal of the American Statistical Association, 1949
    Verify fields before relying on them