arxiv: 2605.04479 · v1 · submitted 2026-05-06 · 💱 q-fin.MF · econ.GN· q-fin.EC

Recognition: unknown

ESG as Priced Crash Insurance: State-Dependent Tail Risk and Deconfounding Evidence

Jiayu Yi, Minxuan Hu, Wenxi Sun, Ziheng Chen

Pith reviewed 2026-05-08 16:36 UTC · model grok-4.3

classification 💱 q-fin.MF econ.GNq-fin.EC

keywords ESGcrash risktail riskdouble machine learningstate-dependent effectspriced insuranceequity crashesdeconfounding

0 comments

The pith

High ESG ratings cut equity crash incidence during systemic drawdowns by shrinking tail losses, acting as priced insurance that costs returns in stable times.

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

The paper sets out to show that ESG operates as state-dependent insurance against equity crashes rather than delivering unconditional performance gains. It defines stress and stable regimes with a drawdown-based truncation rule, then uses double machine learning to separate ESG effects from high-dimensional firm characteristics that would otherwise confound results. The central finding is that high ESG scores lower both the frequency and severity of discrete crashes, but only inside drawdown states and only at the worst quantiles of the return distribution. A reader would care because the result explains mixed ESG performance evidence: the same attribute produces drags when markets are calm yet supplies resilience precisely when losses matter most. The mechanism is asymmetric and distributional, not a simple mean shift.

Core claim

ESG ratings materially reduce the incidence of discrete crash events during systemic drawdowns by attenuating the severity of realized tail losses at the most adverse quantiles. This protection operates as priced insurance that incurs performance drags during stable periods while delivering critical resilience when tail risks materialize. Double machine learning isolates the asymmetric treatment effects by handling complex nuisance parameters and removing high-dimensional confounding between ESG and other firm traits.

What carries the argument

A drawdown-based truncation rule that partitions data into distinct stress and stable economic states, paired with double machine learning that estimates state-dependent treatment effects after removing confounding nuisance parameters.

If this is right

High-ESG portfolios exhibit lower crash frequency and milder tail losses specifically inside drawdown regimes.
The insurance cost appears as lower average returns during stable, non-stress periods.
Standard linear models miss the effect because they do not separate states or remove high-dimensional confounding.
The benefit concentrates at extreme left-tail quantiles rather than shifting the entire return distribution.

Where Pith is reading between the lines

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

Regime-dependent ESG tilts could serve as a dynamic hedging tool for equity portfolios without separate tail-risk instruments.
The same deconfounding and state-partitioning approach could be applied to other firm attributes to test for analogous state-dependent protections.
If the pattern holds in credit or fixed-income data, ESG ratings would influence how markets price systemic tail risk across asset classes.

Load-bearing premise

The drawdown-based truncation rule successfully divides market data into distinct stress and stable states without introducing selection bias, and double machine learning fully removes all confounding between ESG ratings and other firm characteristics.

What would settle it

Finding that high-ESG firms experience the same frequency and severity of crashes as low-ESG firms during a subsequent systemic drawdown, even after applying the same double machine learning deconfounding procedure, would falsify the central claim.

Figures

Figures reproduced from arXiv: 2605.04479 by Jiayu Yi, Minxuan Hu, Wenxi Sun, Ziheng Chen.

**Figure 1.** Figure 1: Market Stress Definition. Notes: The dashed lines mark the 15th-percentile cutoff (−2.82%). Panel (a) shows monthly market returns with stress months highlighted; panel (b) shows the corresponding distribution. effect remains negative deeper in the distribution and is still statistically supported at 𝜏 = 0.20, where the coefficient is −0.015 (95% CI [−0.027, −0.003]). This confirms that the stress classif… view at source ↗

**Figure 2.** Figure 2: DML Treatment Effects of ESG Components by Market Regimes. Notes: This figure illustrates the after-treatment effect estimated via the Double Machine Learning (DML) approach under a Lasso specification. The Agg variable represents the Industry Adjusted Aggregated ESG Score. Components are defined as: E: Environmental Score; S: Social Score; G: Governance Score. Point estimates (𝛽) and 95% confidence interv… view at source ↗

read the original abstract

This research establishes ESG as a state dependent insurance mechanism against equity crashes by addressing the decoupling of unconditional alpha from tail risk resilience. By validating market stress regimes as distinct economic states through a drawdown-based truncation rule, the study demonstrates that high ESG ratings materially reduce the incidence of discrete crash events during systemic drawdowns. To address the selection bias and high-dimensional confounding inherent in traditional linear frameworks, we implement Double Machine Learning as a structural deconfounding layer. Unlike simple predictive modeling, the Double Machine Learning framework utilizes machine learning to handle complex nuisance parameters, allowing us to isolate the asymmetric treatment effects of ESG across different market states. Distributional analysis reveals the underlying mechanism as ESG specifically attenuates the severity of realized tail losses at the most adverse quantiles instead of shifting the entire return distribution. Confirmed by structural estimates, this protection functions as priced insurance that incurs performance drags during stable periods while providing critical resilience when tail risks are most acute.

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.

Desk Editor's Note private letter to a colleague

The paper claims high-ESG stocks cut crash incidence in drawdown states via DML deconfounding and tail attenuation, but the drawdown truncation looks endogenous enough to weaken the insurance framing.

read the letter

The core claim is that ESG acts like priced insurance: it drags returns in normal times but reduces discrete crashes and tail severity during systemic drawdowns. They split data with a drawdown truncation, apply double ML to remove high-dimensional confounding, and show the protection shows up at the worst quantiles rather than shifting the whole distribution. That distributional and state-dependent angle is the main addition over standard ESG alpha or risk papers. The DML step is a reasonable way to handle nuisance parameters in this setting, and the abstract at least tries to separate unconditional effects from conditional tail resilience. Credit for moving past linear regressions on average returns. The stress-test note holds up on the evidence given. Aggregate drawdowns embed the firm-level tail events being studied, so the state indicator can correlate mechanically with the outcome; DML on firm covariates does not fix that. The abstract mentions validation of distinct states but gives no details on the truncation threshold, sensitivity checks, or whether high-ESG firms influence the market index itself. Without those, the asymmetric effect is hard to trust as cleanly identified. Implementation is also thin—no mention of which learners are used for the ML steps, cross-validation, or robustness to alternative state definitions. This is for people working on ESG tail-risk pricing or portfolio insurance who already follow the DML literature in finance. A reader could pull the quantile-attenuation idea and try to tighten the state split, but the current version is too preliminary for direct use. I would send it to peer review with explicit referee guidance to examine the truncation rule and run the obvious robustness tests; the thinking is engaged with the literature even if the identification needs work.

Referee Report

3 major / 1 minor

Summary. The paper claims that ESG ratings function as state-dependent priced insurance against equity crashes: high-ESG firms exhibit materially lower crash incidence during systemic drawdowns identified by a drawdown-based truncation rule, with effects isolated via Double Machine Learning to remove high-dimensional confounding and shown by distributional analysis to concentrate at adverse return quantiles rather than shifting the full distribution, incurring performance costs in stable periods.

Significance. If the identification holds, the results would help reconcile the absence of unconditional ESG alpha with tail-risk resilience, offering a mechanism for state-dependent pricing of ESG characteristics in asset markets. The use of DML for deconfounding and quantile-focused distributional analysis are methodological strengths that could advance empirical work on asymmetric risk exposures.

major comments (3)

[State identification via drawdown truncation] Drawdown truncation rule (abstract and state-identification section): the central claim that this rule cleanly partitions data into exogenous systemic-stress versus normal regimes without selection on ESG or crash outcomes is load-bearing, yet aggregate market returns embed the firm-level tail events under study, risking mechanical correlation between state assignment and the outcome; explicit tests for exogeneity (e.g., placebo partitions or firm-exclusion robustness) are required.
[Double Machine Learning framework] Double Machine Learning implementation (abstract and DML section): the assertion that DML fully removes high-dimensional confounding between ESG and other firm characteristics is central to the deconfounding evidence, but the manuscript supplies no details on nuisance-function estimators, cross-fitting procedure, or sensitivity to hyperparameter choices; without these, it is impossible to verify whether the reported asymmetric tail effects survive alternative nuisance specifications.
[Distributional analysis] Distributional analysis (abstract and results on tail quantiles): the finding that ESG attenuates severity only at the most adverse quantiles (rather than shifting the entire distribution) underpins the insurance interpretation, yet the paper must specify the exact quantile estimation method, its integration with the state-dependent truncation, and whether standard errors account for the two-stage DML procedure.

minor comments (1)

[Abstract] The abstract's phrasing of 'decoupling of unconditional alpha from tail risk resilience' could be clarified with a brief definition of the unconditional benchmark used.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback on our manuscript. The comments highlight important areas for clarification and robustness, and we address each major point below. We plan to incorporate revisions that strengthen the identification, methodological transparency, and replicability of the results.

read point-by-point responses

Referee: [State identification via drawdown truncation] Drawdown truncation rule (abstract and state-identification section): the central claim that this rule cleanly partitions data into exogenous systemic-stress versus normal regimes without selection on ESG or crash outcomes is load-bearing, yet aggregate market returns embed the firm-level tail events under study, risking mechanical correlation between state assignment and the outcome; explicit tests for exogeneity (e.g., placebo partitions or firm-exclusion robustness) are required.

Authors: We agree that the overlap between aggregate market returns and firm-level outcomes introduces a potential mechanical correlation, and that explicit exogeneity tests are warranted. In the revision we will add a dedicated robustness appendix containing: (i) placebo partitions that randomly assign stress dates while preserving the drawdown magnitude distribution, (ii) firm-exclusion checks that recompute the aggregate drawdown after dropping the largest 10 percent of firms by market capitalization, and (iii) an alternative state definition based on VIX thresholds above the 90th percentile. These tests will be reported alongside the baseline results to demonstrate that the state-dependent ESG tail attenuation is not an artifact of the truncation rule. revision: yes
Referee: [Double Machine Learning framework] Double Machine Learning implementation (abstract and DML section): the assertion that DML fully removes high-dimensional confounding between ESG and other firm characteristics is central to the deconfounding evidence, but the manuscript supplies no details on nuisance-function estimators, cross-fitting procedure, or sensitivity to hyperparameter choices; without these, it is impossible to verify whether the reported asymmetric tail effects survive alternative nuisance specifications.

Authors: We acknowledge that the current manuscript lacks sufficient implementation detail for the DML procedure. The revised version will expand the DML section and add an appendix that specifies: the nuisance estimators (random forests for continuous covariates and logistic regression for binary indicators), the cross-fitting scheme (5-fold sample splitting with out-of-fold predictions), and a full set of sensitivity checks that vary key hyperparameters (maximum tree depth, minimum node size, and regularization strength). We will also report the resulting ESG tail coefficients under these alternatives to confirm that the asymmetric attenuation at adverse quantiles remains stable. revision: yes
Referee: [Distributional analysis] Distributional analysis (abstract and results on tail quantiles): the finding that ESG attenuates severity only at the most adverse quantiles (rather than shifting the entire distribution) underpins the insurance interpretation, yet the paper must specify the exact quantile estimation method, its integration with the state-dependent truncation, and whether standard errors account for the two-stage DML procedure.

Authors: We will clarify the distributional analysis in the revised manuscript. The quantile results are obtained via a two-stage procedure that first applies the DML deconfounding step within each regime (stress versus normal) and then estimates conditional quantiles using the deconfounded residuals. We will explicitly state that we employ Koenker’s linear quantile regression at the 5 percent, 10 percent, and 25 percent levels, integrated with the drawdown truncation, and that standard errors are obtained via a bootstrap that resamples both the first-stage nuisance functions and the second-stage quantile regressions to account for the two-stage estimation. These details, together with the bootstrap code outline, will be added to the results section and a methodological appendix. revision: yes

Circularity Check

0 steps flagged

No circularity: derivation uses external data partitions and standard DML estimators without reduction to inputs by construction

full rationale

The paper partitions market regimes via a drawdown-based truncation rule applied to aggregate returns and then applies Double Machine Learning to isolate ESG treatment effects on tail quantiles. No quoted equations or claims equate the estimated asymmetric resilience to a fitted parameter or self-defined outcome; the state indicator is constructed from market-wide returns independent of the firm-level ESG-crash regression, and DML nuisance functions are estimated separately from the target parameters. The abstract and described framework rely on observable return series and off-the-shelf ML deconfounding rather than tautological renaming, self-citation load-bearing, or ansatz smuggling. This leaves the central claim falsifiable against external benchmarks and yields no load-bearing circular steps.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central claim rests on standard econometric assumptions for Double Machine Learning and on the validity of a drawdown-based state definition; no new free parameters, axioms, or invented entities are introduced beyond those implicit in the chosen estimators.

axioms (2)

standard math Double Machine Learning nuisance-parameter estimators are consistent under standard regularity conditions
Invoked to justify isolation of the ESG treatment effect after controlling for high-dimensional confounders.
domain assumption Drawdown-based truncation partitions returns into distinct economic regimes without residual selection bias
Used to define market-stress states; appears in the abstract as the validation step for state-dependent analysis.

pith-pipeline@v0.9.0 · 5474 in / 1287 out tokens · 35879 ms · 2026-05-08T16:36:58.207675+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

15 extracted references · 3 canonical work pages

[1]

Victor Chernozhukov, Denis Chetverikov, Mert Demirer, Esther Du- flo, Christian Hansen, Whitney Newey, and James Robins. 2018. Dou- ble/debiased machine learning for treatment and structural parameters

2018
[2]

Colin Cameron and Douglas L

A. Colin Cameron and Douglas L. Miller. 2015. A Practitioner’s Guide to Cluster-Robust Inference.Journal of Human Resources50, 2 (2015), 317–372. arXiv:https://jhr.uwpress.org/content/50/2/317.full.pdf doi:10. 3368/jhr.50.2.317

2015
[3]

Elizabeth Demers, Jurian Hendrikse, Philip Joos, and Baruch Lev. 2021. ESG did not immunize stocks during the COVID-19 crisis, but invest- ments in intangible assets did.Journal of business finance & accounting 48, 3-4 (2021), 433–462

2021
[4]

2010.Strategic management: A stakeholder approach

R Edward Freeman. 2010.Strategic management: A stakeholder approach. Cambridge university press

2010
[5]

Shihao Gu, Bryan Kelly, and Dacheng Xiu. 2020. Empirical asset pricing via machine learning.The Review of Financial Studies33, 5 (2020), 2223–2273

2020
[6]

Amy P Hutton, Alan J Marcus, and Hassan Tehranian. 2009. Opaque financial reports, R2, and crash risk.Journal of financial Economics94, 1 (2009), 67–86

2009
[7]

Li Jin and Stewart C Myers. 2006. R2 around the world: New theory and new tests.Journal of financial Economics79, 2 (2006), 257–292

2006
[8]

Roger Koenker and Gilbert Bassett Jr. 1978. Regression quantiles.Econo- metrica: journal of the Econometric Society(1978), 33–50

1978
[9]

Hans R. Künsch. 1989. The Jackknife and the Bootstrap for General Stationary Observations.The Annals of Statistics17, 3 (1989), 1217–1241. doi:10.1214/aos/1176347265

work page doi:10.1214/aos/1176347265 1989
[10]

Karl V Lins, Henri Servaes, and Ane Tamayo. 2017. Social capital, trust, and firm performance: The value of corporate social responsibility during the financial crisis.the Journal of Finance72, 4 (2017), 1785– 1824

2017
[11]

MSCI Solutions LLC. 2026. ESG Ratings: Assess companies on their financially relevant sustainability risks and opportunities. https://www. msci.com/data-and-analytics/sustainability-solutions/esg-ratings

2026
[12]

Standard & Poor’s Financial Services LLC. 2026. Compustat North America. https://wrds-web.wharton.upenn.edu/wrds/

2026
[13]

Zhiming Xue, Sichen Zhao, Yalun Qi, Xianling Zeng, and Zihan Yu. 2026. Resilient Routing: Risk-Aware Dynamic Routing in Smart Logistics via Spatiotemporal Graph Learning.arXiv preprint arXiv:2601.13632(2026)

work page arXiv 2026
[14]

Sichen Zhao, Zhiming Xue, Yalun Qi, Xianling Zeng, and Zihan Yu
[15]

Non-Intrusive Graph-Based Bot Detection for E-Commerce Using Inductive Graph Neural Networks.arXiv preprint arXiv:2601.22579 (2026)

work page arXiv 2026