ESG improvers in EUR corporate credit produce a long-short spread of 0.926 basis points a month, with a t of 1.302, according to Table 2. Buying issuers after their sustainability score rises and shorting them after it falls fails as a trade.
The question is where, if anywhere, the signal starts paying.
How the signal is meant to earn money
Henide defines delta ESG momentum as the issuer's month-on-month change in its LSEG ESG score, demeaned across the market each month. A higher score is supposed to capture operational improvement. Such improvement reduces left-tail risk, and bondholders, facing capped upside and full downside, should value that reduction differently from equity holders.
The score alone cannot distinguish a changed business from better disclosure. Henide therefore filters the delta for credibility. He uses two proxies: the LSEG ESG Controversy score and the issuer's presence on public institutional exclusion lists compiled by the Financial Exclusions Tracker.
The sample covers the non-financial corporate sub-sample of Markit iBoxx EUR Overall at month-end from 31 December 2014 to 30 April 2025. Requiring LSEG ESG Score coverage leaves 158,206 line items, comprising 3,380 unique bonds and 434 unique issuers across 114 monthly return observations. The benchmark applies a minimum cut-off of EUR 500 million to EUR-issued corporate bonds. During the study window, 178 of the 434 issuers appeared on a public exclusion list maintained by at least one unique investor. Green, Social or Sustainability flags applied to 242 of the 3,380 bonds.
Returns require care. The paper proxies them with the month-on-month change in option-adjusted spread, reversing the level order so that spread compression counts as positive. Its figures in "basis points per month" are spread changes rather than total or excess returns. No transaction costs are applied to the quintile strategy.
The abstract says the strategy "earns up to 88 basis points monthly". Table 4 creates a problem for that claim. Its note also describes the magnitudes as basis points per month, yet the largest marginal effect is 0.877 bps. The figures differ by a factor of 100 within the paper.
The signal is sparse. Henide acknowledges that annual LSEG review cycles leave relatively few non-zero ESG momentum observations. Most identification in the monthly framework therefore comes from rating update events, which occur only a small number of times per issuer each year.
Exclusions reverse the stated story
The main conditional finding comes from interactions between ESG momentum and exclusion tiers. ESG times Excluded 1-4 is 0.752 (p<0.001), while ESG times Excluded 5+ is 0.939 (p=0.002). Across 154,825 bond-months, the base ESG momentum coefficient is -0.062, with a standard error of 0.071 and p=0.384.
The corresponding marginal effect without exclusions is -0.062, and its confidence interval crosses zero. One to four exclusions lifts the effect to 0.690. Five or more raises it to 0.877.
The largest premium belongs to issuers that the greatest number of investors have publicly refused to own.
The abstract describes credibility as "low ESG controversy and minimal institutional investor exclusions". The conclusion goes further, treating the absence of institutional investor exclusions as a precondition for reward. Yet the exclusion coefficients fit a scrutiny interpretation. The paper eventually adopts that interpretation too, arguing that "genuine sustainability improvements carry the greatest information value for issuers whose prior conduct has attracted the most scrutiny".
Henide addresses the possible mechanical explanation. Excluded issuers may already trade at wider spreads, allowing score improvements to coincide with compression that would have occurred anyway. He cites controls for OAS level and bid-ask spread in response, although Table 3 contains no OAS-level term in its specification.
He also says the effect gathers in the upper tail instead of appearing symmetrically across quantiles, citing Table 6. Yet Table 6 reports no exclusion-tier interaction at any quantile. The rows cover the ESG base effect, ESG times BBB, ESG times Subordinated, ESG times Sub times OAS and ESG times Sub times LowControv.
The controversy measure comes closer to the advertised mechanism. ESG times Low Controversy is 0.368 (p=0.001), and the Low Controversy dummy itself contributes 1.778. Table 5 divides the sample at the panel median of 73.08. Its note says the metric uses an inverse scale, meaning higher scores correspond to lower controversy, while the dummy marks issuers with below-median controversy.
Table 6 then applies a threshold below 25. According to its note, this threshold identifies "subordinated issuers with the greatest controversy exposure within that segment". The results text describes the identical group as "the most controversy-free subordinated issuers". Adjacent tables assign opposite meanings to the same variable name.
Tail insurance, then a very thin subordinated sample
At the tenth percentile, the base ESG momentum effect is +0.46 bps (p<0.01). It falls to -0.10 at the median (p<0.01). The seventieth and ninetieth percentiles show -0.06 and -0.02, respectively, and both are insignificant.
The sign changes as returns improve. ESG momentum behaves more like insurance than a persistent return source, paying during bad states and surrendering a little around the middle. The estimate uses all 154,825 bond-month observations.
Koenker and Machado pseudo-R-squared measures quantile-regression fit. It is 0.026 at Q0.1, 0.001 at the median and 0.054 at Q0.9. Explanatory power at the top decile is fifty times its level at the centre. Henide interprets this concentration in the extremes as evidence for the credibility barrier framework.
The subordinated estimates deserve far less confidence. At Q0.9, the triple interaction among ESG momentum, subordinated status and OAS reaches 143.84, with a standard error of 27.30. The same term at Q0.1 is -15.00, with a standard error of 215.28.
Those coefficients come from 1,082 subordinated bond-months, just 0.70% of the panel. Henide acknowledges the problem, writing that the magnitudes are "sensitive to individual bond-month observations in this relatively thin segment". He also states that ESG momentum times subordinated times low-controversy is estimated in a separate two-interaction specification because the full model cannot be identified in that sub-segment.
ESG times Subordinated times Green cannot be estimated at all because fewer than 20 observations qualify. Henide converts the absence into an argument: instruments that most need credibility are precisely those that issuers do not bring to market under a green label. Even so, subordinated debt leads the abstract's second sentence.
Three inconsistencies remain. The results section gives the long-short spread as 0.50 bps with a 5.31 standard deviation, whereas Table 2 reports 0.926 and 7.592. In the descriptive table, the subordinated dummy has a mean of 0.000 and a standard deviation of 0.007, figures that cannot be reconciled with 1,082 observations among 158,205. The third is the 88 bps headline beside Table 4's 0.877 bps.
Regulation gives the cleaner result
SFDR took effect in March 2021. Henide first estimates a difference-in-differences using Developed Europe domicile as treatment. He then adds green-label status in a triple difference. A further version weights observations by inverse propensity and includes half-year time fixed effects.
The broad estimate survives these changes: Developed Europe-domiciled bonds gain +0.32 bps per month after March 2021, with p<0.001. Under the stricter specification, the additional label-specific effect shrinks to 0.08 bps and p=0.421.
Compliance flows therefore moved eligible bonds according to domicile. Once composition and time trends are absorbed, the green label earns nothing separately. In the main panel, the green dummy is already negative at -2.837 (p=0.001). Henide's policy conclusion follows from this pattern: diffuse pricing instead of label-concentrated pricing gives issuers less incentive to spend on label integrity.
The treatment variable also weakens the claimed identification. Developed Europe domicile is a geographic division, one that overlaps with the 2021 increase in inflation expectations and rate volatility discussed by Henide. Sector composition also differs between EU and non-EU EUR issuers.
Henide confronts both issues. A placebo dated January 2020, immediately before COVID, produces no effect. He also argues that geographic identification makes industry-by-time fixed effects the wrong control because the split is geographic rather than sectoral. The placebo carries information. It leaves the sector-composition concern unresolved.
Were the exclusion lists point in time?
We ran nothing.
The strategy requires issuer-level ESG score revisions plus public exclusion lists mapped onto EUR cash bonds. We hold neither those inputs nor a universe retaining bond-level duration and seniority.
One issue determines whether the exclusion finding deserves treatment beyond an in-sample result. The bond-month exclusion flag may begin when the exclusion was eventually disclosed, or when investors could first have known it. The former treatment would introduce look-ahead into the time-varying dummies. We did not find any discussion in the paper of vintage or as-of handling for the Financial Exclusions Tracker data.
The shape of the result remains. ESG score changes in credit appear state-contingent, delivering +0.46 bps at Q0.1 versus -0.10 at the median, where pseudo-R-squared is 0.001. Henide separates state-contingent pricing from a factor premium, and the unconditional long-short spread of 0.926 bps a month resolves the factor question.
Crossing the bid-ask costs more. The panel's average bid-ask is 0.442 price points. At its mean modified duration of 5.456-year, that equates to roughly 8 bps of spread, several times the signal.
We have previously examined ESG research in which the ratings input carried every result (Synthetic ESG Ratings Drive Every Result). Henide uses genuine vendor ratings. The weak point has shifted to the conditioning set.