Green finance, ESG & sustainable finance: mainstream in form, contested in substance.

Academic interest in sustainable finance has followed an explosive trajectory — a bibliometric read of 9,218 publications indexed in Web of Science and Scopus (2012–2025) shows interest accelerating sharply after 2018, with compound annual growth near 55% in some sub-fields. The inflection coincides with the European Green Deal, the first supranational green-bond issuance, and central banks turning their attention to climate as a financial risk. But volume came with a conceptual shift: from arguing that financing green projects is the right thing to do, toward pricing sustainability as a risk-management imperative embedded in core regulation.

The stakes are not abstract. The global ESG market exceeded $41 trillion in AUM in 2024; green and sustainable bond issuance has compounded into a multi-trillion-dollar market; and a disclosure stack — TCFD, ISSB, CSRD, the EU Taxonomy — is reshaping the information environment for every listed company. This review covers five domains and the cross-cutting tension that binds them: a field that is mainstream in form, but whose measurement foundation remains fragile.

The review spans 2018–2026 across SSRN, arXiv (econ, q-fin), Web of Science, Scopus, PubMed Central, and official output from the ECB, BIS, IMF, the European Commission and NGFS, with emphasis on 2023–2026. Selection favoured formal economic modelling, rigorous identification (matched samples, propensity scoring, event studies, synthetic controls, panel fixed effects), or systematic-review method — not normative advocacy. Five domains were chosen for their direct bearing on financial risk, asset pricing and prudential policy.

The decisive structural split is between use-of-proceeds instruments (the label commits the issuer to where money goes) and performance-linked instruments (the coupon itself adjusts on KPIs). The latter is more incentive-compatible but contractually fragile — verification, and the temptation to set unambitious baselines. Across both, issuer commitment and credible third-party verification are the primary determinants of credibility and price.

The greenium — the yield discount investors accept for green bonds versus otherwise-equivalent conventional debt — is the most intensively studied question in sustainable fixed income. A matched-sample comparison at issuance puts it near 16 bp on average. But the average conceals a two-tier structure: the discount roughly doubles to ~32 bp when the issuer's environmental score sits in the top tercile, and certification plus climate-uncertainty periods amplify it further. The finding — "it better be good, it better be green" — says sophisticated investors price the credibility of the claim, not the label, consistent with signalling under information asymmetry.

Regional heterogeneity is large. Europe shows robust, consistent greeniums backed by the Taxonomy and deep demand; China's market shows diluted, inconsistent discounts under weaker certification and state-driven price signals; in Korea, a 274-pair matched study (2020–2025) confirms a greenium that differs by issuer type. A two-factor structural model rationalises the spread: orthogonal cash-flow risk (the issuer) and effectiveness risk (the green project) can generate positive or negative greeniums depending on their correlation — so conflicting empirics are a feature, not an anomaly. Separately, wavelet network analysis finds green bonds and green equities behave as two largely independent asset classes — green debt does not hedge climate equity risk inside a sustainable portfolio.

ESG ratings from commercial providers (MSCI, Sustainalytics, Bloomberg, Refinitiv, ISS) disagree so sharply for the same firm that their cross-provider correlation runs 0.30–0.60 — against 0.99 for Moody's vs S&P credit ratings of equivalent instruments. That is a level incompatible with the informational role ratings are meant to play. The Berg–Kölbel–Rigobon decomposition attributes the gap to three sources, with the first two doing most of the work:

• Scope divergence — raters measure different attributes of the same concept.
• Measurement divergence — raters use different proxies for the same attribute.
• Weighting divergence — raters weight the same indicators differently (smaller than commonly assumed).

The consequences are material. Rating disagreement raises information asymmetry for lenders: firms with high inter-rater ESG dispersion obtain fewer newly originated bank loans, controlling for fundamentals. For equity investors, provider choice becomes a methodological confound — performance attribution shifts with the data vendor, contaminating the entire ESG–performance literature. Calibrated benchmarking from verifiable data, not self-reported surveys, is the necessary condition for coherence.

The chronic limits of traditional ratings have catalysed an AI-augmented literature. LLMs deployed via Retrieval-Augmented Generation extract structured sustainability signals from disclosures and external sources with far less lag than periodic rater cycles. In controlled comparisons, portfolios built on AI-derived ESG scores post higher mean returns and superior Sharpe ratios than both AI-based low-ESG and traditionally rated ESG portfolios — with lower downside risk and smaller maximum drawdowns under stress. Regression confirms AI-derived scores hold a stronger association with excess returns than conventional ESG metrics.

Greenwashing detection is the high-value application. Zero-shot classification of TCFD-aligned disclosures can flag whether a bank's climate narrative is substantive or boilerplate without manual labelling, scored across the four TCFD pillars (governance, strategy, risk management, metrics and targets). Graph-based representations of LLM-extracted data then test consistency across reporting periods — the core of detecting greenwashing at scale rather than case by case.

The TCFD frame — now inside ISSB's IFRS S2 and the EU's CSRD/ESRS — splits climate risk into physical (acute event damage; chronic multi-decade impairment) and transition (policy/carbon pricing, technology/stranded assets, market and reputational channels). Finance has leaned on transition risk because it is partially observable through policy signals; physical risk needs asset-level climate downscaling, a data problem only partly solved. A six-factor model augmenting Fama–French with a pollutant-minus-green (PMG) factor finds a significant carbon premium across regions, stronger in small caps.

But the headline result is a null. A quarterly panel of 238 MSCI Europe ESG Leaders (2018–2024) finds no stable linear carbon return premium — the coefficient is indistinguishable from zero, with only nonlinear specifications revealing curvature in the middle of the carbon distribution. Controversy risk, driven by negative ESG incidents, is the most robust pricing signal — not carbon intensity per se. NYC's Local Law 97 supplies a clean natural experiment: the 2019 announcement produced insignificant abnormal returns, but enforcement onset in January 2024 coincided with meaningful negative returns for REITs with NYC office exposure — policy uncertainty resolving at implementation. On stranded assets, capital-requirement work shows that if high-carbon assets are merely 13% riskier than current risk weights reflect, a buffer concentrated in systemic banks would protect the system — the analytic basis for the green-supporting / dirty-penalising factor (GSF/DPF) debate.

The NGFS scenario framework — orderly transition (early, predictable policy), disorderly (late or abrupt tightening), and hot house world (insufficient action, severe physical risk) — is now the standard for supervisory exercises at the ECB, Bank of England, Banque de France and Bank of Canada. Climate risk has shifted from optional disclosure to a regulated component of prudential oversight.

The ECB's first climate stress test (2022) found aggregate short-term transition costs manageable for European banks, but physical-risk costs rising disproportionately in the hot house world over a 30-year horizon. The exposure distribution is highly skewed: transition risk concentrates in a few banks with large carbon-intensive loan books, physical risk concentrates geographically (Southern Europe, coastal assets) — so systemic risk is lower in aggregate but more severe for specific institutions, a microprudential-versus-macroprudential distinction. The binding constraint is structural: climate scenarios run 20–30 years, while capital planning runs 1–3, and macrofinancial models estimated on short cycles cannot extrapolate cleanly. The extrapolation problem is fundamental, not merely technical.

The disclosure landscape evolved in three overlapping waves. TCFD (2017) set the conceptual frame and reached over 4,000 organisations voluntarily. ISSB (IFRS S1/S2, 2023) operationalised it into a global baseline. The EU's CSRD (2025) goes furthest: double materiality — disclosing both financial materiality to investors and impact materiality to society — under ESRS, extending mandatory reporting to roughly 4× the companies of the prior NFRD regime, with auditor assurance and Scope 3. The political-economy tension then surfaced: the European Commission proposed cutting ESRS requirements by up to 50%, a retreat reflecting SME compliance-cost resistance.

The EU Taxonomy classifies activities via technical screening (DNSH, minimum safeguards, contribution to one of six objectives), covering activities responsible for up to 80% of EU emissions. But a formal welfare analysis is cautious: Taxonomy-based lending preferences generate income losses from productivity differences while emissions don't change at a fixed EU ETS cap — and if brown production relocates outside the EU rather than decarbonising, worldwide emissions can even rise, since CBAM covers only specific sectors at the border. This "carbon leakage via capital allocation" critique implies the Taxonomy's effectiveness is conditional on the parallel carbon-pricing regime, not independently powerful. The bank-level Green Asset Ratio has, predictably, revealed limited Taxonomy alignment of current loan books.

A bibliometric analysis of 652 papers (2017–2025) tracks greenwashing research from a green-marketing/legitimacy phase into an AI-governance, quantitative-detection phase. The governance response has crystallised around three instruments — mandatory Taxonomy alignment (EU GBS), mandatory external review at instrument level, and CSRD issuer-level assured disclosure. Meanwhile the 2022–2025 politicisation wave in the US saw BlackRock, Vanguard and State Street withdraw from or moderate climate coalitions (Climate Action 100+, Net Zero Asset Managers). Global flows haven't reversed, but the result is a geographic bifurcation — EU integration under mandate, US retreat under political and legal pressure — with implications for cross-jurisdiction cost-of-capital differentials. On the frontier, the TNFD (400+ committed organisations) extends disclosure to nature, and blended finance is the primary catalytic mechanism against the $4–5tn/year gap concentrated in EMDEs.

Sustainable finance is mainstream in form: the instruments are real, the regulatory architecture is being built, the flows are macroeconomically consequential. But the measurement foundation is fragile, the causal evidence on additionality is weaker than the marketing implies, and US politics injects genuine uncertainty into the world's largest capital market. The most consequential question for 2026–2030 is whether the field can deliver environmental outcomes at scale under contested measurement and fragmented regulation — or whether integration into core financial regulation demands a deeper resolution of the measurement crisis that underpins its entire architecture.