Abstract
Global policies pursuing environmental goals have proliferated over recent decades. They have, on the one hand, spurred green branding—a strategic repositioning by firms to align with ESG values—and, on the other, accelerated ESG investment flows that are reshaping finance. In this context, mortgage lending has emerged as a lever to mobilise additional private capital for home renovation and to lift the energy performance of the housing stock—an urgent priority given that most buildings that will be in use over the next 25 years in Europe are already built. Green mortgages bring banks into the campaign for greener buildings at critical decision points, when owners plan refurbishments and efficiency upgrades. For lenders, these products can mean lower credit risk (via reduced utility bills and greater borrower disposable income), enhanced collateral quality (a potential “green premium”), and reduced exposure to “brown discounts” that may render inefficient properties less attractive to buyers and investors.
To attract borrowers, however, banks must communicate effectively—often beginning with digital marketing pages and online product announcements. This paper maps the current European landscape of banks supporting green mortgages, examines how lenders define the “green” element of a mortgage in the absence of a binding EU-level definition, and identifies the green-branding strategies most commonly used to promote these loans. It also analyses the affective framing of bank–homebuyer communication, probing the boundary between legitimate persuasion and potentially manipulative advertising, and assesses whether emerging practices risk a new form of greenwashing.
Methodologically, we construct two corpora of texts—green and standard mortgage offers—collected through targeted sitographical research of major and smaller European retail banks, and we apply complementary text-mining strategies. Topics are identified via Latent Dirichlet Allocation. Sentiment polarity is first measured using lexicon-based approaches (AFINN and Bing) and then triangulated with a finance-specific pipeline that combines the R package sentometrics with the Loughran–McDonald lexicon, allowing us to capture risk-, uncertainty-, and litigation-laden language that generic lexicons miss. We then cross-validate polarity and uncertainty estimates with a domain transformer (FinBERT/FiBERT), which provides probabilistic classifications for positive/neutral/negative tone and finance-salient categories (e.g., risk and uncertainty). Emotional profiling is performed using the NRC Emotion Lexicon to classify discrete emotions such as trust, fear, and anticipation alongside global polarity. Lexical frequencies, distributions, and co-occurrence structures are analysed with widyr and visualised with igraph/ggraph; distinctive terms are detected using log-likelihood ratios with quanteda.textstats. Inference combines Welch’s t-tests and Cohen’s d for continuous sentiment scores, χ² and binomial tests for categorical emotion proportions, and Kolmogorov–Smirnov together with Jensen–Shannon divergence for distributional differences.
To ensure identification is not driven by artefacts, we run placebo tests—including label permutation between green and standard pages and token shuffling that preserves frequent n-grams—and a broad suite of robustness checks across preprocessing choices (stemming/lemmatisation and boilerplate removal), language and bank-size strata, sampling schemes, alternative lexicons, and alternative model specifications; we also implement leave-one-bank-out re-estimation. The main results remain stable across these perturbations and under cross-validation between sentometrics/Loughran–McDonald and FinBERT/FiBERT.
The comparative analysis reveals a consistent pattern of semantic and affective differentiation. The green-mortgage corpus exhibits a statistically significant over-representation of positively valenced language and of emotions such as anticipation, joy, and trust, together with attenuated salience of cost and risk disclosures. Topic modelling shows that energy efficiency, future value, and upgrades tend to co-occur with performative promise language rather than detailed eligibility criteria or quantified savings. Co-occurrence networks display hubs organised around aspirational frames (“upgrade,” “future value,” “comfort”), while regulatory and risk nodes (APR, variable-rate risk, fees) are peripheralised. These patterns are visible in lexicon-based metrics, reproduced by sentometrics with the Loughran–McDonald dictionary, and confirmed by FinBERT/FiBERT through higher probabilities of positive tone and lower probabilities of uncertainty/risk for the green corpus. Taken together, the evidence suggests an affective-salience imbalance that can plausibly reduce critical risk assessment by consumers at the margin, especially where benefits are credence-based and intertemporally distant.
The contribution to law and economics is twofold. Substantively, we translate the linguistic construction of “green” into measurable variables—affective balance, salience-adjusted disclosure, and finance-specific uncertainty—thereby operationalising informational asymmetries and attention biases within models of bounded rationality and welfare assessment under EU consumer-protection law. Conceptually, we offer a law-and-language bridge: marketing claims are treated as speech acts whose illocutionary force (assurance, promise) yields detectable perlocutionary effects on sentiment and attention; our pipeline makes these effects computable and testable, providing an evidentiary basis for enforcement and for a calibrated, “paternalism-lite” regulatory stance.
A policy recommendation follows directly: the EU should introduce a standardised Green Mortgage Information Sheet (GMIS), harmonised with the Taxonomy and accompanied by a brief, numeracy-first disclosure of eligibility criteria, expected energy-savings ranges with uncertainty bands, and salient price/risk items. Supervisors should require an ex-ante affective-balance check—using benchmarked sentometrics and FinBERT/FiBERT scores—to ensure that positive framing does not overwhelm risk salience beyond a rebuttable threshold; materials failing the check would trigger enhanced disclosure or redesign. Such a regime aligns environmental objectives with consumer protection and reduces the risk that green branding drifts into misleading presentation or greenwashing.
To attract borrowers, however, banks must communicate effectively—often beginning with digital marketing pages and online product announcements. This paper maps the current European landscape of banks supporting green mortgages, examines how lenders define the “green” element of a mortgage in the absence of a binding EU-level definition, and identifies the green-branding strategies most commonly used to promote these loans. It also analyses the affective framing of bank–homebuyer communication, probing the boundary between legitimate persuasion and potentially manipulative advertising, and assesses whether emerging practices risk a new form of greenwashing.
Methodologically, we construct two corpora of texts—green and standard mortgage offers—collected through targeted sitographical research of major and smaller European retail banks, and we apply complementary text-mining strategies. Topics are identified via Latent Dirichlet Allocation. Sentiment polarity is first measured using lexicon-based approaches (AFINN and Bing) and then triangulated with a finance-specific pipeline that combines the R package sentometrics with the Loughran–McDonald lexicon, allowing us to capture risk-, uncertainty-, and litigation-laden language that generic lexicons miss. We then cross-validate polarity and uncertainty estimates with a domain transformer (FinBERT/FiBERT), which provides probabilistic classifications for positive/neutral/negative tone and finance-salient categories (e.g., risk and uncertainty). Emotional profiling is performed using the NRC Emotion Lexicon to classify discrete emotions such as trust, fear, and anticipation alongside global polarity. Lexical frequencies, distributions, and co-occurrence structures are analysed with widyr and visualised with igraph/ggraph; distinctive terms are detected using log-likelihood ratios with quanteda.textstats. Inference combines Welch’s t-tests and Cohen’s d for continuous sentiment scores, χ² and binomial tests for categorical emotion proportions, and Kolmogorov–Smirnov together with Jensen–Shannon divergence for distributional differences.
To ensure identification is not driven by artefacts, we run placebo tests—including label permutation between green and standard pages and token shuffling that preserves frequent n-grams—and a broad suite of robustness checks across preprocessing choices (stemming/lemmatisation and boilerplate removal), language and bank-size strata, sampling schemes, alternative lexicons, and alternative model specifications; we also implement leave-one-bank-out re-estimation. The main results remain stable across these perturbations and under cross-validation between sentometrics/Loughran–McDonald and FinBERT/FiBERT.
The comparative analysis reveals a consistent pattern of semantic and affective differentiation. The green-mortgage corpus exhibits a statistically significant over-representation of positively valenced language and of emotions such as anticipation, joy, and trust, together with attenuated salience of cost and risk disclosures. Topic modelling shows that energy efficiency, future value, and upgrades tend to co-occur with performative promise language rather than detailed eligibility criteria or quantified savings. Co-occurrence networks display hubs organised around aspirational frames (“upgrade,” “future value,” “comfort”), while regulatory and risk nodes (APR, variable-rate risk, fees) are peripheralised. These patterns are visible in lexicon-based metrics, reproduced by sentometrics with the Loughran–McDonald dictionary, and confirmed by FinBERT/FiBERT through higher probabilities of positive tone and lower probabilities of uncertainty/risk for the green corpus. Taken together, the evidence suggests an affective-salience imbalance that can plausibly reduce critical risk assessment by consumers at the margin, especially where benefits are credence-based and intertemporally distant.
The contribution to law and economics is twofold. Substantively, we translate the linguistic construction of “green” into measurable variables—affective balance, salience-adjusted disclosure, and finance-specific uncertainty—thereby operationalising informational asymmetries and attention biases within models of bounded rationality and welfare assessment under EU consumer-protection law. Conceptually, we offer a law-and-language bridge: marketing claims are treated as speech acts whose illocutionary force (assurance, promise) yields detectable perlocutionary effects on sentiment and attention; our pipeline makes these effects computable and testable, providing an evidentiary basis for enforcement and for a calibrated, “paternalism-lite” regulatory stance.
A policy recommendation follows directly: the EU should introduce a standardised Green Mortgage Information Sheet (GMIS), harmonised with the Taxonomy and accompanied by a brief, numeracy-first disclosure of eligibility criteria, expected energy-savings ranges with uncertainty bands, and salient price/risk items. Supervisors should require an ex-ante affective-balance check—using benchmarked sentometrics and FinBERT/FiBERT scores—to ensure that positive framing does not overwhelm risk salience beyond a rebuttable threshold; materials failing the check would trigger enhanced disclosure or redesign. Such a regime aligns environmental objectives with consumer protection and reduces the risk that green branding drifts into misleading presentation or greenwashing.