技能 数据科学 金融数学理论数值实验指导

金融数学理论数值实验指导

v20260724
mathfin-data-analysis
本指南指导作者如何在数学金融论文中进行严格、可复现的数值实验。其核心原则是将数值工作作为支撑理论证明的工具(如收敛率、误差界),而非进行实证数据分析或市场数据拟合。重点在于数值方法的严谨性、过程的可追溯性和对理论的支撑性。
获取技能
138 次下载
概览

Numerical Experiments (mathfin-data-analysis)

Note on framing

This is a theory-first journal. Mathematical Finance explicitly states that **numerical
experiments are welcome only when accompanied by a rigorous analysis** supporting the
theoretical developments, and that **routine application of computational methods to financial
data will not be considered**. So "data analysis" here is not empirical estimation — it is
numerical work that illustrates or stress-tests a theorem. This skill is deliberately lighter
than its empirical-journal counterpart.

When to trigger

  • You want to add simulations or a numerical scheme to a proof-based paper
  • A referee may ask whether your theorem "does anything" beyond existence
  • You need to show convergence, accuracy, or qualitative behavior predicted by the theory

How to keep numerics journal-appropriate

  1. Tie every experiment to a result. Each figure/table should illustrate a specific

theorem, proposition, or rate (e.g., "Monte Carlo error decays at the proven $O(n^{-1/2})$
rate", "the free boundary matches the smooth-fit characterization").

  1. State the method precisely. Discretization scheme (Euler–Maruyama, Milstein, PDE

finite-difference/finite-element), step sizes, number of paths, variance reduction,
truncation of the domain — enough that the experiment is reproducible.

  1. Report error, not just output. Where the theory gives a rate or bound, show the

empirical rate against it; show convergence as the grid refines.

  1. Choose parameters with financial meaning (volatilities, maturities, strikes) so the

illustration speaks to the modelling problem.

  1. Keep numerics subordinate. They support the theory; they are never the contribution.

Do not let a numerical section grow into a stand-alone empirical study.

Reproducibility (light but real)

  • Pin software/library versions; set and report random seeds for any Monte Carlo.
  • Make illustrative code reproducible; consider archiving it (Zenodo/GitHub) and citing it.
  • Include a Data Availability Statement even if no external data are used (see

mathfin-replication-and-data-policy).

Matching scheme to result type

Result being illustratedNatural schemeWhat the exhibit must report
Strong/weak SDE convergence rateEuler–Maruyama or Milstein with halving stepslog–log error slope against the proven order
BSDE well-posedness or rateBackward Euler / least-squares Monte Carlo / deep BSDE solverterminal error and driver residual across grids
Optimal stopping / free boundaryBinomial tree or PDE variational-inequality solverboundary location against the smooth-fit characterization
Rough-volatility approximationHybrid scheme for fractional kernels; Markovian liftimplied-vol skew slope against the proven power law
Duality gap = 0Primal candidate and dual bound computed independentlygap shrinking as the discretization refines
Mean-field limitN-player simulation vs. McKean–Vlasov solverdistance to the limit decaying in N at the stated rate

Worked micro-example: convergence exhibit for a rough-volatility paper

Suppose Theorem 3.2 proves that a Markovian multi-factor approximation of a rough volatility
model converges at a rate governed by the Hurst parameter H. The journal-appropriate exhibit:
simulate both models with the same Brownian increments, plot the implied-volatility error
against the number of factors on log axes, draw the theoretical slope as a reference line, and
caption with the scheme, step size, path count, seed, and the theorem number. What would NOT
fit: calibrating the approximation to index-option data and reporting fit quality — that turns
an illustration into the empirical study the journal screens out.

Pre-submission numerics audit

  • Every exhibit names the theorem, proposition, or rate it illustrates — no orphan plots.
  • The observed rate is computed (regression slope), not eyeballed, and stated next to the

proven one.

  • Degenerate sanity cases (zero volatility, Black–Scholes limit, H → 1/2) reproduce known

closed forms before the general runs are trusted.

  • The numerical section would survive deletion: the theorems stand alone without it.

Anti-patterns

  • A numerical study with no theorem behind it (out of scope for this journal).
  • Plots with no error/convergence analysis where the theory promises a rate.
  • Unstated scheme, step size, or path count — irreproducible.
  • Calibrating to real market data and presenting it as the paper's result.

Output format

【Experiment】what it illustrates (which theorem/rate)
【Method】scheme + step/paths + variance reduction
【Error reported】empirical vs. theoretical rate/bound
【Parameters】financial values used
【Reproducibility】seeds + versions + code location
【Next step】mathfin-tables-figures
信息
分类 数据科学
名称 mathfin-data-analysis
版本 v20260724
大小 5.22KB
更新时间 2026-07-28
语言