Paper: arXiv 2510.15458

Authors: Gabriele Visentin, Patrick Cheridito

Abstract

In this paper, we show that interventionally robust optimization problems in causal models are continuous under the \(G\)-causal Wasserstein distance, but may be discontinuous under the standard Wasserstein distance. This highlights the importance of using generative models that respect the causal structure when augmenting data for such tasks. To this end, we propose a new normalizing flow architecture that satisfies a universal approximation property for causal structural models and can be efficiently trained to minimize the \(G\)-causal Wasserstein distance. Empirically, we demonstrate that our model outperforms standard (non-causal) generative models in data augmentation for causal regression and mean-variance portfolio optimization in causal factor models.

Complexity vs Empirical Score

  • Math Complexity: 9.5/10
  • Empirical Rigor: 6.0/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper is heavily theoretical with dense mathematical proofs (e.g., continuity under G-causal Wasserstein distance, universal approximation properties), warranting a high math score. It provides empirical validation in finance (causal regression, portfolio optimization) with clear comparisons to baselines, but the implementation details and data specifics are less emphasized than the theoretical framework.

Research Flowchart

  flowchart TD
  A["Research Goal:<br>Robust Optimization in Causal Models"] --> B["Key Methodology:<br>Interventionally Robust Optimization<br>via G-Causal Wasserstein Distance"]
  B --> C["Computational Process:<br>Proposed G-Causal Normalizing Flow<br>for Efficient Training"]
  C --> D["Data Input:<br>Causal Factor Models &<br>Interventional Data"]
  D --> E["Comparative Analysis:<br>Standard vs. Causal<br>Generative Models"]
  E --> F["Key Findings:<br>1. Robustness continuous under G-Causal Wasserstein<br>2. Outperforms non-causal models in<br>causal regression & portfolio optimization"]

Evidence Card

What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card

Proposes

The paper proposes a new normalizing flow architecture called G-causal normalizing flows that universally approximates causal structural models.

Source sentence
we propose a new normalizing flow architecture that satisfies a universal approximation property for causal structural models

Assumes 2

  • The DAG G and target set T satisfy that G quotiented by the partition {T} ∪{{i}, i ∈V \ T} is a DAG. model
    Source sentence
    Assumption 3.1. The DAG G = (V, E) and the target set T ⊆V are such that G quotiened by the partition {T} ∪{{i}, i ∈V \ T} is a DAG.
  • The DAG G is a finite directed acyclic graph with vertices sorted so that if (i, j) ∈E then i < j. model
    Source sentence
    We assume throughout that G = (V, E) is a given directed acyclic graph (DAG) with a finite index set V = {1, . . . , d}, which we assume, without loss of generality, to be sorted (i.e. (i, j) ∈E, then i < j).

Data

  • linear Gaussian SCM with bipartite DAG synthetic unknown
    100 stocks and 20 factors
    Source sentence
    Gaussian SCM, with bipartite DAG G with partition {T, V \T} and random uniform coefficients in (−1, 1), and γ is a given risk aversion parameter. The target variables XT represent stock returns, while XV \T are market factors or trading signals.

Costs

Unclear - sources disagree The platform classifier and the paper-text extraction give different answers, so no value is asserted.

Platform classifier: Not modeled

Paper text extraction: Not applicable

Finds 3

  • G-causal normalizing flows outperform standard non-causal generative models in data augmentation for causal regression and mean‑variance portfolio optimization. performance
    Source sentence
    We show empirically that G-causal normalizing flows outperform non-causal generative models (such as variational auto-encoders, standard normalizing flows, and nearest‑neighbor KDE) when used to perform generative data augmentation in two empirical setups: causal regression and mean‑variance portfolio optimization in c
  • The Sharpe ratio of the G‑causal portfolio remains robust across a wide range of intervention strengths, while non‑causal portfolios deteriorate rapidly. performance
    Source sentence
    Fig. 9 and Fig. 10 show that the Sharpe ratio of the G-causal portfolio is indeed robust to a wide range of interventions, while the performance of non-causal portfolios deteriorates rapidly, starting from the least favorable quantiles.
  • Causal optimization problems are continuous under the G‑causal Wasserstein distance but may be discontinuous under the standard Wasserstein distance. theoretical
    Source sentence
    We prove that causal optimization problems (i.e. problems in which optimizers must be functions of the state variables’ parents in the causal DAG G) are continuous under the G-causal Wasserstein distance, but may be discontinuous under the standard Wasserstein distance.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 23 related papers). Coverage is partial, so this is neither replication nor refutation.

Code

No repository by the authors found in the paper.

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.

Cited References (34)

  1. Scaling Learning Algorithms Towards AI (2007)
  2. A Fast Learning Algorithm for Deep Belief Nets (2006)
  3. Deep learning (2016)
  4. Time-Causal VAE: Robust Financial Time Series Generator — → 2411.02947
  5. Set-Valued Analysis (2009)
  6. Adapted Wasserstein distances and stability in mathematical finance (2020)
  7. The probability of backtest overfitting (2017)
  8. Measure theory (2007)
and 26 more
  1. Triangular transformations of measures (2005)
  2. A brief review of hypernetworks in deep learning (2024)
  3. Distributionally robust learning (2020)
  4. A comprehensive survey for generative data augmentation (2024)
  5. Optimal transport and Wasserstein distances for causal models (2025)
  6. Computational methods for adapted optimal transport (2024)
  7. Quantitative stability of regularized optimal transport and convergence of sinkhorn’s algorithm (2022)
  8. Real Analysis: Modern Techniques and their Applications (1999)
  9. Neural autoregressive flows (2018)
  10. Probability Essentials (2004)
  11. Distributionally robust optimization (2025)
  12. Graphical Models (1996)
  13. Multilayer feedforward networks with a nonpolynomial activation function can approximate any function (1993)
  14. Quantitative Risk Management: Concepts, Techniques and Tools (Revised Edition) (2015)
  15. Methods and tools for causal discovery and causal inference (2022)
  16. Elements of Causal Inference: Foundations and Learning Algorithms (2017)
  17. Restricted Structural Equation Models for Causal Inference (2012)
  18. A distance for multistage stochastic optimization models (2012)
  19. Multistage Stochastic Optimization (2014)
  20. Variational Analysis (1998)
  21. Invariant Models for Causal Transfer Learning (2018)
  22. Spline Functions: Basic Theory (2007)
  23. Lectures on Stochastic Programming: Modeling and Theory (2021)
  24. Cot-gan: Generating sequential data via causal optimal transport (2020)
  25. A survey on causal discovery: theory and practice (2022)
  26. Toward understanding generative data augmentation (2023)

1 of 34 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.