pith. sign in

arxiv: 2304.01207 · v4 · submitted 2023-03-24 · 💱 q-fin.CP · math.PR· q-fin.RM

A Multilevel Stochastic Approximation Algorithm for Value-at-Risk and Expected Shortfall Estimation

St\'ephane Cr\'epey (LPSM (UMR\_8001)) , Noufel Frikha (CES) , Azar Louzi (LPSM (UMR\_8001)) This is my paper

Pith reviewed 2026-05-24 09:47 UTC · model grok-4.3

classification 💱 q-fin.CP math.PRq-fin.RM
keywords value-at-riskexpected shortfallmultilevel stochastic approximationnested simulationstochastic approximationcomplexity analysisrisk estimation
0
0 comments X

The pith

A multilevel stochastic approximation scheme achieves near-optimal complexity for estimating Value-at-Risk and Expected Shortfall in nested simulation settings.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces a multilevel stochastic approximation algorithm for computing Value-at-Risk and Expected Shortfall when these quantities require nested Monte Carlo simulations due to conditional risk factors. Standard nested stochastic approximation has complexity of order ε^{-3} for accuracy ε. The new MLSA method improves this to ε^{-2-δ} for VaR, where δ depends on loss integrability, and to ε^{-2} |ln ε|^2 for ES. This is demonstrated through theoretical analysis and numerical studies showing reduced execution time for given error rates.

Core claim

The MLSA algorithm attains an optimal complexity of the order ε^{-2-δ} for VaR estimation and ε^{-2}|ln ε|^2 for ES estimation, where δ∈(0,1) depends on the integrability degree of the loss, compared to the ε^{-3} complexity of standard nested stochastic approximation.

What carries the argument

The multilevel stochastic approximation (MLSA) scheme, which combines stochastic approximation with multilevel Monte Carlo ideas to handle biased innovations in nested problems.

If this is right

  • VaR estimation complexity improves from cubic to sub-cubic in the reciprocal of accuracy.
  • ES estimation achieves nearly quadratic complexity up to a logarithmic factor.
  • The method regains performance lost due to the nested nature of the problem.
  • Joint evolution of error rate and execution time shows significant gains in numerical tests.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The approach may extend to other nested stochastic approximation problems in finance or statistics.
  • Similar multilevel techniques could apply to higher-dimensional risk measures.
  • The dependence of δ on integrability suggests tailoring the method to specific loss distributions.

Load-bearing premise

The loss random variable satisfies moment conditions that allow a positive δ less than 1 in the complexity bound for VaR.

What would settle it

A simulation study measuring the actual runtime scaling with decreasing ε and comparing to the predicted rates of ε^{-2-δ} and ε^{-2} log squared would confirm or refute the complexity claims.

Figures

Figures reproduced from arXiv: 2304.01207 by Azar Louzi (LPSM (UMR\_8001)), Noufel Frikha (CES), St\'ephane Cr\'epey (LPSM (UMR\_8001)).

Figure 1
Figure 1. Figure 1: Centered and rescaled risk measures for a bias parameter h tending to 0. The left panel plot suggests that, asymptotically as H ∋ h ↓ 0, the quantities ξ h ⋆ − ξ 0 ⋆ and χ h ⋆ − χ 0 ⋆ are linear in h. The right panel plot strengthens this observation, demonstrating that the quantities h −1 (ξ h ⋆ − ξ 0 ⋆ ) and h −1 (χ h ⋆ − χ 0 ⋆ ) are approximately constant in a neighborhood of 0. This checks empirically … view at source ↗
Figure 2
Figure 2. Figure 2: Performance comparison of Algorithms 1, 2 and 3. 15 [PITH_FULL_IMAGE:figures/full_fig_p015_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Performance comparison of Algorithms 1, 2 and 3. SA scheme VaR estimation ES estimation RMSE Accuracy RMSE Accuracy Nested SA (Alg. 2) −3.75 −3.00 −3.11 −3.00 Multilevel SA (Alg. 3) −4.50 −2.93 −2.35 −2.51 SA (Alg. 1) −1.57 −2.00 −1.99 −2.00 [PITH_FULL_IMAGE:figures/full_fig_p019_3.png] view at source ↗
read the original abstract

We propose a multilevel stochastic approximation (MLSA) scheme for the computation of the value-at-risk (VaR) and expected shortfall (ES) of a financial loss, which can only be computed via simulations conditionally on the realisation of future risk factors. Thus the problem of estimating its VaR and ES is nested in nature and can be viewed as an instance of stochastic approximation problems with biased innovations. In this framework, for a prescribed accuracy $\varepsilon$, the optimal complexity of a nested stochastic approximation algorithm is shown to be of the order $\varepsilon^{-3}$. To estimate the VaR, our MLSA algorithm attains an optimal complexity of the order $\varepsilon^{-2-\delta}$, where $\delta\in(0,1)$ is some parameter depending on the integrability degree of the loss, while to estimate the ES, the algorithm achieves an optimal complexity of the order $\varepsilon^{-2}|\ln{\varepsilon}|^2$. Numerical studies of the joint evolution of the error rate and the execution time demonstrate how our MLSA algorithm regains a significant amount of the performance lost due to the nested nature of the problem.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper proposes a multilevel stochastic approximation (MLSA) algorithm for nested estimation of Value-at-Risk (VaR) and Expected Shortfall (ES) under simulation-based conditional loss distributions. It establishes that standard nested stochastic approximation has complexity O(ε^{-3}), while the proposed MLSA attains O(ε^{-2-δ}) for VaR (with δ ∈ (0,1) depending on loss integrability) and O(ε^{-2} |ln ε|^2) for ES, supported by numerical studies of error versus runtime.

Significance. If the complexity bounds hold under the stated integrability conditions, the result would represent a meaningful advance in computational finance by mitigating the cost of nested risk-measure estimation, with the numerical experiments providing concrete evidence of practical gains over the baseline.

major comments (2)
  1. [Abstract] Abstract: the headline VaR rate O(ε^{-2-δ}) is load-bearing for the central claim yet rests on moment conditions that produce δ > 0. The manuscript must state the precise moment threshold (e.g., E[|L|^{p}] for which p yields δ > 0) and verify it for standard loss distributions such as log-normal or heavy-tailed portfolios; if only second moments hold, the multilevel bias-variance analysis typically forces δ = 0 and the complexity reverts to the nested baseline.
  2. [Theoretical analysis] Theoretical analysis (presumably the main theorem deriving the rates): the bias-variance decomposition and the dependence of δ on integrability must be fully derived with explicit regularity conditions; without these derivations the claimed optimality cannot be assessed and the numerical studies alone do not confirm the rates.
minor comments (1)
  1. [Numerical studies] Numerical studies section: include explicit error-bar details, sample sizes, and the precise loss distributions used so that the observed runtime-error scaling can be reproduced and compared against the theoretical predictions.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments on our manuscript. We address each major comment point-by-point below, agreeing where clarification is needed and proposing targeted revisions.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the headline VaR rate O(ε^{-2-δ}) is load-bearing for the central claim yet rests on moment conditions that produce δ > 0. The manuscript must state the precise moment threshold (e.g., E[|L|^{p}] for which p yields δ > 0) and verify it for standard loss distributions such as log-normal or heavy-tailed portfolios; if only second moments hold, the multilevel bias-variance analysis typically forces δ = 0 and the complexity reverts to the nested baseline.

    Authors: We agree that the moment threshold should be stated explicitly. The parameter δ > 0 requires E[|L|^{2+η}] < ∞ for some η > 0 (with δ = η/(2+η) in the bias-variance analysis). Log-normal losses satisfy this for any η since all moments exist. Heavy-tailed portfolios satisfy it whenever moments strictly above order 2 are finite. We will revise the abstract and add a short remark with this threshold and verification. Under the paper's stated integrability the rate does not revert to O(ε^{-3}). revision: yes

  2. Referee: [Theoretical analysis] Theoretical analysis (presumably the main theorem deriving the rates): the bias-variance decomposition and the dependence of δ on integrability must be fully derived with explicit regularity conditions; without these derivations the claimed optimality cannot be assessed and the numerical studies alone do not confirm the rates.

    Authors: Theorem 3.1 and its proof in Section 4 already contain the bias-variance decomposition under Assumptions 2.1–2.3, which encode the integrability of L and show how multilevel corrections yield the δ-dependent rate. We will revise the theorem statement to display the explicit mapping from moment index to δ and expand the proof sketch with the precise regularity conditions used in the variance bound. This will make the derivation fully self-contained. revision: partial

Circularity Check

0 steps flagged

No circularity; complexity bounds derived from multilevel analysis

full rationale

The paper states that nested SA has complexity ε^{-3} and derives improved rates ε^{-2-δ} (VaR) and ε^{-2}|ln ε|^2 (ES) for the MLSA scheme under stated integrability assumptions on the loss. These rates are presented as consequences of the bias-variance analysis of the multilevel estimator; no equation reduces a claimed prediction to a fitted parameter by construction, no load-bearing uniqueness theorem is imported via self-citation, and no ansatz is smuggled. The moment conditions are explicit assumptions, not self-referential definitions. The derivation chain is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claims rest on standard moment assumptions for the loss random variable and on the general theory of stochastic approximation with biased innovations; no new entities are introduced.

free parameters (1)
  • δ
    Integrability parameter in (0,1) that appears in the VaR complexity bound; its value is determined by the loss distribution rather than fitted to data in the abstract.
axioms (1)
  • domain assumption The loss admits moments of sufficiently high order to guarantee a positive δ < 1
    Invoked to obtain the ε^{-2-δ} rate stated in the abstract.

pith-pipeline@v0.9.0 · 5763 in / 1337 out tokens · 22753 ms · 2026-05-24T09:47:44.873027+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Adaptive Multilevel Stochastic Approximation of the Value-at-Risk

    q-fin.RM 2024-08 unverdicted novelty 7.0

    An adaptive multilevel stochastic approximation scheme for Value-at-Risk computation achieves complexity O(ε^{-2} |ln ε|^{5/2}) by selecting inner samples adaptively at each level.

  2. Asymptotic Error Analysis of Multilevel Stochastic Approximations for the Value-at-Risk and Expected Shortfall

    q-fin.RM 2023-11 unverdicted novelty 6.0

    Establishes central limit theorems for the renormalized estimation errors of nested and multilevel stochastic approximation algorithms for VaR and ES, including averaged versions, with numerical illustration.

Reference graph

Works this paper leans on

25 extracted references · 25 canonical work pages · cited by 2 Pith papers

  1. [1]

    On the coherence of expected shortfall

    C. Acerbi and D. Tasche. “On the coherence of expected shortfall”. In:Journal of Banking & Finance26.7 (2002), pp. 1487–1503

    work page 2002

  2. [2]

    XVAmetricsforCCPoptimization

    C.Albanese,Y.Armenti,andS.Crépey.“XVAmetricsforCCPoptimization”.In: Statistics & Risk Modeling37.1-2 (2020), pp. 25–53

    work page 2020

  3. [3]

    Computing VaR and CVaR using stochastic approx- imation and adaptive unconstrained importance sampling

    O. Bardou, N. Frikha, and G. Pagès. “Computing VaR and CVaR using stochastic approx- imation and adaptive unconstrained importance sampling”. In:Monte Carlo Methods and Applications 15.3 (2009), pp. 173–210

    work page 2009

  4. [4]

    CVaR hedging using quantization-based stochastic approximation algorithm

    O. Bardou, N. Frikha, and G. Pagès. “CVaR hedging using quantization-based stochastic approximation algorithm”. In:Mathematical Finance26.1 (2016), pp. 184–229

    work page 2016

  5. [5]

    Recursive computation of value-at-risk and condi- tionalvalue-at-riskusingMCandQMC

    O. Bardou, N. Frikha, and G. Pagès. “Recursive computation of value-at-risk and condi- tionalvalue-at-riskusingMCandQMC”.In: Monte Carlo and Quasi-Monte Carlo Methods. Springer Berlin Heidelberg, 2009, pp. 193–208

    work page 2009

  6. [6]

    Stochastic ap- proximation schemes for economic capital and risk margin computations

    D. Barrera, S. Crépey, B. Diallo, G. Fort, E. Gobet, and U. Stazhynski. “Stochastic ap- proximation schemes for economic capital and risk margin computations”. In:ESAIM: Proceedings and Surveys65 (2019), pp. 182–218

    work page 2019

  7. [7]

    Multilevel Monte Carlo methods and lower–upper bounds in initial margin computations

    F. Bourgey, S. De Marco, E. Gobet, and A. Zhou. “Multilevel Monte Carlo methods and lower–upper bounds in initial margin computations”. In:Monte Carlo Methods and Appli- cations 26.2 (2020), pp. 131–161

    work page 2020

  8. [8]

    Risk estimation via regression

    M. Broadie, Y. Du, and C. Moallemi. “Risk estimation via regression”. In:Operations Research 63.5 (2015), pp. 1077–1097

    work page 2015

  9. [9]

    Non asymptotic controls on a recursive superquantile approxi- mation

    M. Costa and S. Gadat. “Non asymptotic controls on a recursive superquantile approxi- mation”. In:Electronic Journal of Statistics15.2 (2021), pp. 4718–4769

    work page 2021

  10. [10]

    Crépey, N

    S. Crépey, N. Frikha, A. Louzi, and G. Pagès. Asymptotic error analysis of multilevel stochastic approximations for the value-at-risk and expected shortfall. 2023. arXiv: 2311. 15333 [q-fin.RM]

    work page 2023

  11. [11]

    General multilevel adaptations for stochastic approx- imation algorithms of Robbins-Monro and Polyak-Ruppert type

    S. Dereich and T. Müller-Gronbach. “General multilevel adaptations for stochastic approx- imation algorithms of Robbins-Monro and Polyak-Ruppert type”. In:Numerische Mathe- matik 142.2 (2019), pp. 279–328

    work page 2019

  12. [12]

    Föllmer and A

    H. Föllmer and A. Schied.Convex Risk Measures. John Wiley & Sons, Ltd, 2010

    work page 2010

  13. [13]

    Multi-level stochastic approximation algorithms

    N. Frikha. “Multi-level stochastic approximation algorithms”. In:The Annals of Applied Probability 26.2 (2016), pp. 933–985

    work page 2016

  14. [14]

    Shortfall risk minimization in discrete time financial market models

    N. Frikha. “Shortfall risk minimization in discrete time financial market models”. In:SIAM Journal on Financial Mathematics5.1 (2014), pp. 384–414

    work page 2014

  15. [15]

    A multi-step Richardson–Romberg extrapolation method for stochastic approximation

    N. Frikha and L. Huang. “A multi-step Richardson–Romberg extrapolation method for stochastic approximation”. In:Stochastic Processes and their Applications125.11 (2015), pp. 4066–4101

    work page 2015

  16. [16]

    Concentration bounds for stochastic approximations

    N. Frikha and S. Menozzi. “Concentration bounds for stochastic approximations”. In:Elec- tronic Communications in Probability17 (2012), pp. 1–15

    work page 2012

  17. [17]

    Multilevel Monte Carlo path simulation

    M. B. Giles. “Multilevel Monte Carlo path simulation”. In:Operations Research56.3 (2008), pp. 607–617

    work page 2008

  18. [18]

    Multilevel nested simulation for efficient risk estimation

    M. B. Giles and A.-L. Haji-Ali. “Multilevel nested simulation for efficient risk estimation”. In: SIAM/ASA Journal on Uncertainty Quantification7.2 (2019), pp. 497–525

    work page 2019

  19. [19]

    M. B. Giles, A.-L. Haji-Ali, and J. Spence.Efficient risk estimation for the credit valuation adjustment. 2024. arXiv:2301.05886 [q-fin.CP]. 45

    work page arXiv 2024

  20. [20]

    Antithetic multilevel Monte Carlo estimation for multidi- mensional SDEs

    M. B. Giles and L. Szpruch. “Antithetic multilevel Monte Carlo estimation for multidi- mensional SDEs”. In:Monte Carlo and Quasi-Monte Carlo Methods 2012. Springer Berlin Heidelberg, 2013, pp. 367–384

    work page 2012

  21. [21]

    Weak error for nested multilevel Monte Carlo

    D. Giorgi, V. Lemaire, and G. Pagès. “Weak error for nested multilevel Monte Carlo”. In: Methodology and Computing in Applied Probability22 (3 2020), pp. 1325–1348

    work page 2020

  22. [22]

    Nested simulation in portfolio risk measurement

    M. B. Gordy and S. Juneja. “Nested simulation in portfolio risk measurement”. In:Man- agement Science56.10 (2010), pp. 1833–1848

    work page 2010

  23. [23]

    Adaptive multilevel Monte Carlo for probabilities

    A.-L. Haji-Ali, J. Spence, and A. L. Teckentrup. “Adaptive multilevel Monte Carlo for probabilities”. In:SIAM Journal on Numerical Analysis60.4 (2022), pp. 2125–2149

    work page 2022

  24. [24]

    Optimization of conditional value-at-risk

    R. T. Rockafellar and S. Uryasev. “Optimization of conditional value-at-risk”. In:Journal of Risk 2.3 (2000), pp. 21–41

    work page 2000

  25. [25]

    Efficient risk estimation via nested multilevel quasi-Monte Carlo simulation

    Z. Xu, Z. He, and X. Wang. “Efficient risk estimation via nested multilevel quasi-Monte Carlo simulation”. In:Journal of Computational and Applied Mathematics 443 (2024), p. 115745. 46

    work page 2024