The study's entire negative premium is 0.2241 basis points, far too little to offer a trade. Its own tests cannot distinguish the result from zero. The Wilcoxon p is 0.2399, while the two-sample t-test p is 0.4297. Hachenberg reports no statistically significant negative premium in the full sample and writes that "we cannot accept our first hypothesis".
A small market under an ICMA label
Social bonds are use-of-proceeds paper. Issuers commit the proceeds to projects with a social purpose, then label the debt against the ICMA Social Bond Principles. Published in 2017 and modelled on the green bond equivalents, those principles remain voluntary industry guidelines.
No EU Social Taxonomy exists behind them. A report was originally due in 2021, a Platform on Sustainable Finance document followed in 2022, and no decision has been made on whether to build the taxonomy at all. The paper calls the resulting risk socialwashing.
Even within sustainable debt, the social bond market is small. Hachenberg's filters identify 545 eligible social bonds beside 56,987 comparable active conventional bonds, under 1% of that universe. The paper places both within a green, social and sustainability (GSS) segment worth roughly 3% of the total bond market. Cumulative GSS issuance reached USD 6.8tn by the end of 2025.
The 280 bonds left after matching shape every result that follows. Of these, 205 (73%) have a rating from Moody's, S&P or Fitch. Their mean rating on the paper's numerical scale is 4.1, roughly Aa3/AA-. This is high-grade paper. Asia contributes 144 bonds and Europe 118, leaving about 93.5% of the sample in those regions. The 75 unrated bonds come from 13 issuers, and 72 of them are Korean or Japanese. Payment rank is similarly concentrated: 169 are senior unsecured (over 60%), while 95 are secured (about 34%). ESG assurance is almost universal, covering 276 of 280 bonds, more than 98%.
The measurement
The pricing case follows the argument used for green paper. Demand for labelled sustainable debt exceeds supply, investors accept a slightly lower required yield, and tighter funding repays the issuer's cost of obtaining a second-party opinion. Hachenberg lays out that mechanism while acknowledging that the greenium literature she reviews remains divided over whether any discount exists.
The study builds matched triplets. Each social bond gets two conventional bonds from the same issuer, one shorter and one longer. Rating, payment rank and currency must match, and every bond must be fixed-coupon and plain vanilla. Linear interpolation between the two conventional i-spreads produces a synthetic spread at the social bond's maturity. The daily differential equals the social spread minus that synthetic spread.
Daily trading-day data covers 1 October 2023 to 30 September 2024. The finished sample contains 280 social bonds and 560 conventional bonds, issued by 70 issuers across 16 countries and 11 currencies. That produces 50,465 social-bond observations and 100,930 conventional ones. The attrition is severe. The initial 1,983 candidates shrink to 280 after 1,383 fall below the EUR 150m issue size floor and another 216 lack two matchable comparables.
Mean social i-spread is 32.6345 bps, against 32.8586 bps for the synthetic bonds. Standard deviations are 45.2266 and 44.9328 respectively. Mean issue size is EUR 1.07bn for the social bonds and EUR 1.24bn for the conventional matches.
Tighter by a fifth of a basis point. Smaller by EUR 170m.
Why do spread and yield studies disagree?
Hachenberg identifies the divide directly. Research using i-spreads or asset-swap spreads tends to find almost nothing. Intonti and co-authors estimate a mean of -2.32 bps and a median of -0.37 bp. Pop and co-authors obtain a significant differential of -6.73 and -6.37 bps in exactly one of five matched groups. This paper lands at -0.22 bp.
Yield studies produce much larger estimates, sometimes with the opposite sign. Baldi and Ferri find a new-issue yield "socium" of 47 bps. García-Escobar and co-authors estimate an EU social premium of 13.89 bp, compared with a 4.22 bp greenium. Using daily bid yields, Torricelli and Pellati report a positive social premium of 1.242 bp. The estimates span +1.2 to -47 bps, largely according to whether the researcher first removes the interest rate component. Anyone pricing labelled paper in credit should distrust a quoted yield differential on that basis.
Fifty thousand rows, seventy issuers
The abstract claims that ESG assurance, region and timing affect the negative premium. Only one of those three withstands scrutiny.
ESG assurance has the largest coefficient in the bond-specific panels. It is -4.7872 bps at p<0.001 in the population-averaged model, then falls to -3.7730 after adding currency and rank controls. Yet almost no sample variation identifies it. More than 98% of the bonds have an assurance provider, leaving a contrast based on 4 bonds from 3 issuers. Hachenberg acknowledges the problem: "we must be cautious in its interpretation." The result does not belong in the abstract.
The European issuer dummy survives. Across the full sample, its coefficient ranges from -2.5637 to -2.6913 bps at p<0.01. Restrict the panel to the 36,887 observations with complete one-year price series and the estimate remains -1.8606 at p<0.05.
Industry effects fare worse. Government-related coefficients range from -5.18 to -6.55 bps in the full panel and are significant in six of eight specifications. Corporate coefficients run from -4.95 to -7.39 bps. Once the reduced sample is used, those estimates range from -1.31 to 1.93 and none remains significant. Hachenberg attributes the difference to outliers in the larger sample. Fair enough. The reduced sample deserves the citation.
Timing provides the weakest claim. Quarterly dummy coefficients range from -0.1660 to +0.3599 bp. The p<0.001 stars appear solely in the random-effects columns. Population-averaged columns report nearly identical coefficients of -0.1658, 0.1605 and 0.3597, without stars. In the reduced sample, q2 reverses from +0.1602 in the full panel to -0.3550 and becomes insignificant. q3 retains its stars at +0.1426 with p<0.001.
Hachenberg concedes this weakness in the conclusion. The premium is "visible, yet insignificant, in 2023", while the differential across cuts is "very small, ranging from 0.2 to 1.4 bp, and is rarely significant". She interprets the drift as capital leaving sustainable investments, citing ESG fund outflows and geopolitical pressure. The explanation may be right. Twelve months of daily data cannot separate it from the 2023 to 2024 rate cycle, and a 0.36 bp quarterly shift asks a single year to represent a cycle.
Arithmetic creates the larger problem. Statistical significance comes from pooling 50,465 daily rows over a cross-section of 280 bonds and 70 issuers. Government, sovereign or supranational issuers account for 178 of the 280 bonds (63.5%), while Asia and Europe together supply around 93.5% of the sample. Clustering by ISIN leaves issuer-level dependence untouched. The Wilcoxon and t-test results also disagree over the relevant subsamples. AA is significant only under the t-test; A, BBB and NR appear only under Wilcoxon. A-rated and financial social bonds trade wider than their conventional matches.
Nothing to trade, something to retain
An average differential of 0.22 bp against a 32.6 bp mean spread equals 0.7% of the level. The regressions contain no bid-ask or transaction-cost measure. Issue size is their sole liquidity proxy, and it is insignificant in every specification of Table 8. We read the 0.22 bp differential as smaller than any plausible bid-ask for bonds of this kind.
We could not test any of it. Doing so requires security-level social labels, ICMA alignment flags, payment rank and daily i-spreads for individual cash bonds from the same issuer. Our instruments cover US stocks, ETFs, crypto, listed futures and EOD options. Bond ETFs cannot serve as substitutes because aggregated holdings remove the same-issuer matching that gives the differential its meaning.
Hachenberg is explicit that the two literatures measure different things. Baldi and Ferri and García-Escobar price new issues; this paper examines secondary-market i-spreads. Its result therefore limits what investors should expect from labelled paper already trading. It settles nothing about the new-issue concession. The finding also supplies a second credit-market data point beside what we found in EUR credit. A multi-year i-spread panel that preserves the European dummy while controlling for liquidity would change my view of the regional effect. Another single-year cross-section would leave it unchanged.