Motivation
Does brown pay? The carbon premium debate revisited
The carbon premium — the excess return of brown firms over green counterparts — is one of the most contested empirical questions in climate finance. Bolton & Kacperczyk (2021, 2023) find a positive premium; Bauer et al. (2022), Zhang (2025), and Aswani et al. (2024) challenge it. This paper undertakes a systematic analysis across methodologies, data sources, and geographies, using more than 3,500 US firms and more than 10,000 firms across 90 countries from 2007 to 2023 — including COVID-19, the energy crisis, and the Ukraine war.
1
Main finding: green outperforms
Robust evidence of a negative carbon premium in the US and globally — green firms tend to outperform brown firms when using emissions levels as the greenness measure.
2
Results are methodologically fragile
Green outperformance is primarily driven by vendor-estimated emissions and disappears once emissions are scaled by revenues or proper data lags are applied to align with investor information sets.
3
Data quality matters enormously
Extensive cleaning of LSEG Workspace data reveals systematic errors. Quantitative estimates change substantially with cleaned vs. uncleaned data, though qualitative conclusions are similar.
Data & sample
Scope 1 and 2 CO₂e emissions, LSEG Workspace, 2007–2023
The analysis uses firm-level financial and CO₂e emissions data from LSEG Workspace. Scope 1 and scope 2 emissions serve as the primary environmental measure; scope 3 is excluded due to disclosure inconsistencies and double-counting concerns. The distinction between reported and vendor-estimated emissions is central to all results.
US sample
3,592 firms · Feb 2007 – Jan 2023
Panel regressions and portfolio analyses with double-clustered standard errors (firm × month). Industry- and time-fixed effects throughout. Green outperformance is most pronounced here, with monthly GMB alphas of 1.02%–1.30% after FF5 and momentum adjustment.
Global sample
10,516 firms · 90 countries · 831,314 obs.
Country-fixed effects added throughout. Green outperformance persists but is 40%–60% smaller than in the US. Portfolio alphas turn insignificant once the US is excluded, suggesting the phenomenon is largely US-driven.
Vendor-estimated vs. reported emissions
LSEG Workspace estimates emissions for non-disclosing firms via three sequential methods (CO₂, Energy, Median). Only half of covered firms self-report. The paper shows that in panel regressions, all green outperformance is driven by vendor-estimated emissions — restricting the sample to firms that disclose their own emissions yields no significant effect. This raises measurement error and endogeneity concerns.
Empirical framework
Panel regressions, portfolio analysis, and the climate concern test
Panel regression (baseline and with industry fixed effects)
Reti,t = α + β Emissionsi,t + δ Controlsi,t + μt + λindustry + εi,t
Run at the firm-month level. Emissions is either log emissions levels or emissions intensity (emissions scaled by revenues). Controls follow Bolton & Kacperczyk (2021): leverage, sales growth, momentum, volatility, book-to-market, investment/assets, log size, ROE. Standard errors double-clustered at firm and month levels.
GMB portfolio spread — Fama–French alpha
RetGMBt = α + β1RMRFt + β2SMBt + β3HMLt + β4RMWt + β5CMWt + β6UMDt + εt
Monthly value-weighted Green-Minus-Brown spread: long greenest quintile, short brownest quintile. The intercept α captures the return differential unexplained by the Fama–French 5 factors and momentum. Newey–West standard errors throughout.
Key result 1 — main finding
Green firms outperform — but only when sorting on unscaled emissions levels
Cumulative returns of the greenest versus brownest quintile portfolios show persistent green outperformance from 2009 onwards when sorting on emissions levels. Sorting on emissions intensity largely eliminates this gap, both in the US and globally.
Figure 1. Cumulative returns (%) for green and brown portfolios in the US sample (Feb 2007–Jan 2023). Panel (a) sorts on emissions levels — green outperformance is pronounced from mid-2009, reaching ~450% by 2021 vs. ~230% for brown. Panel (b) sorts on emissions intensity — the gap nearly disappears, both lines ending near 280%, highlighting the decisive role of scaling.
Panel regression: green outperformance, US
A firm with vendor-estimated log emissions at the 20th percentile earns on average 2.18 percentage points more per month than a comparable firm at the 80th percentile (with industry fixed effects). For the combined sample, outperformance ranges from 1.20% to 1.30% points per month prior to lagging.
Portfolio analysis: alpha after risk adjustment
After controlling for FF5 factors and momentum, the GMB alpha for the combined US sample ranges from 1.02% to 1.30% per month. The monthly Sharpe ratio of the GMB portfolio (0.234) exceeds that of the market portfolio (0.160) over the full sample period.
Global vs. US: weaker outside the US
Panel regression coefficients in the non-US sample are 50%–70% smaller than in the US and portfolio alphas turn insignificant once the US is excluded — consistent with higher climate awareness and regulatory pressure being concentrated in the US market over the sample period.
Climate concerns do not explain it
The interaction coefficient ρ on Emissions × UMC is statistically insignificant in all specifications. Green outperformance cannot be attributed to unexpected surges in media climate change concerns — contrary to the mechanism proposed by Pastor et al. (2021).
Key result 2 — methodological sensitivity
Four dimensions determine the sign and significance of the carbon premium
Reported vs. vendor-estimated emissions
In panel regressions, the negative carbon premium is entirely driven by vendor-estimated emissions. Restricting to self-reporting firms yields no significant effect. This raises concerns about measurement error and endogeneity when mixing reported and estimated data — a concern shared by Aswani et al. (2024), though they find a positive premium for vendor data.
Emissions levels vs. emissions intensity
Scaling emissions by revenues purges virtually all firm-size information and eliminates green outperformance in all specifications. Log revenues alone explain 64%+ of variation in log emissions levels, implying strong multicollinearity when log market capitalisation is also included as a control variable.
Temporal alignment of emissions and accounting data
Emissions are typically reported 10–12 months after fiscal year-end, later than accounting data. Lagging emissions by 6 or 10 months (and accounting data by 6 months) attenuates green outperformance in panel regressions — though portfolio results are more robust to lagging, preserving significance for the combined sample.
Panel regression vs. portfolio analysis
Portfolio analyses show green outperformance even for reported emissions in the US (alpha 1.02%–1.30%), unlike panel regressions. The methods differ in weighting, treatment of firm controls, and sensitivity to multicollinearity — panel regressions are more sensitive to scaling and vendor-data choice, portfolio analyses to the geographic scope.
Why it matters
Implications for investors, data providers, and regulators
Asset managers & ESG investors
Green outperformance, if genuine, would justify tilting toward low-emission firms. But the conclusion depends critically on whether vendor-estimated data is included and emissions are left unscaled. Investors relying on different data vendors or scaling choices may reach opposite conclusions about the sign of the carbon premium.
Data providers & academic researchers
The paper documents systematic errors in LSEG Workspace emissions data — wrong measurement units, partial restatements, inconsistencies between scope 1+2 sums and reported totals. Cross-validation across vendors and rigorous cleaning are essential prerequisites before using emissions data in return regressions.
Regulators & disclosure standards
The fragility of results to whether firms self-report argues strongly for mandatory, standardised emissions disclosure. The negative premium found for vendor-estimated data may reflect measurement error rather than true carbon risk pricing — a direct concern for regulatory frameworks that rely on emissions data to assess climate-related financial risk.
Caution on the existing literature
The sign of the carbon premium depends on the data vendor, whether emissions are scaled, the time period, the geography, and the empirical model. Any single specification should be interpreted with care. The broader debate between Bolton & Kacperczyk and their challengers — including Aswani et al. and Zhang — remains empirically unresolved.