The Resource-Based View (RBV) assumes optimal and enduring organisational performance results from resources that are not only valuable but also rare and impossible to imitate or replace ( Asif et al., 2023). These conditions are met by intellectual systems, qualified human capital, and capabilities related to sustainability due to their tacit, path-dependent, and socially complex character, and, as a result, they are imitatively costly for competitors ( Nadeem et al., 2024). Such non-financial assets of financial institutions complement pipeline-based innovation, credit risk analytics, regulatory compliance and service quality, which, in turn, enable them to become long-run competent ( Buallay et al., 2022). It is important to note that disclosure does not ascertain the presence of intangible resources, but rather externalises their presence, formation, and strategic action, and that firms with better intangible endowments are in a better position to achieve a positive sustainability outcome ( Menicucci and Paolucci, 2023).

These managerial information advantages exacerbate information asymmetry between capital suppliers and insiders, as substantial value in intellectual, human, and natural capital is not visible to non-experts ( Aevoae et al., 2023). The signaling theory is one of the theories, according to which voluntary disclosure is an expensive and plausible type of latent quality information. Lessens asymmetry by reporting systematically on workforce development, innovation infrastructure, green lending, and environmental risk management by financial institutions, indicating preparation, maturity in governance, and long-term orientation ( Di Tommaso and Thornton, 2023). The salience of signaling is higher in new markets where the institutional enforcement is less, and the strength of the information environment is weak, which enhances the marginal value of credible disclosure ( Rahman et al., 2022; Sobhani et al., 2022)

The above-mentioned views may be supplemented by the stakeholder theory, which explains that companies report non-financial information to ensure that the heterogeneous claims of regulators, investors, workers, customers and civil society are met ( Aboud and Diab, 2022). The legitimacy theory, in its turn, implies that organisations request disclosed information to grant them an assurance that they are meeting the established social norms to maintain the right to operate, which has been intensified by banks in the climate-related supervisory imperative ( Galletta and Mazzù, 2023; Bhuiyan et al., 2024). Combined, RBV explains the performance logic, the communication logic highlighted by signaling theory, and the social-accountability logic of intangible capital disclosure in the context of stakeholder/legitimacy theories.

2.4.1.1. Empirical Literature.

Intellectual Capital Disclosure and Sustainability

The list of positive correlations between IC and firm-level results is quite impressive. The study of European integrated reporters by Salvi et al. (2022) et indicates that the quality of IC disclosure is very high, which is linked to the high market value and reduced cost of capital. Vitolla et al. (2022) transfer this observation to a larger sample of the world, and Buallay et al. (2022) demonstrate that the factors of IC play an essential role in a positive influence on the performance of the banks involved in the operations and financial performance in the context of the MENA markets. Literature that explicitly relates sustainability performance to the IC disclosure is not very rich, nontheless Ali et al. (2024) and Nadeem et al. (2024) demonstrate the positive impact of the IC on the ESG performance, in particular, in knowledge-heavy industries. It is a conceptually sound relationship that is not yet well developed in the banking sector, where intangible inputs play a larger role in value creation, particularly in the non-Western world.

2.4.2 Human capital Disclosure and Sustainability.

Empirical studies on human capital disclosure have been growing in the literature, indicating that. employee training, diversity and transparency in welfare are expected to be linked to greater productivity and better governance ratings, as shown by Cavicchi and Vagnoni (2023) and Lambooy et al. (2022). Furthermore, Ali et al. (2024) and Rahman and Akhter (2023) establish a positive correlation between workforce-dependent disclosure and ESG composite scores, as well as a decrease in involuntary employee turnover. In a bibliometric synthesis of the post-pandemic literature, Mahmud et al. (2022) affirm these trends. This evidence is to a great extent available in developed markets; similar evidence in emerging financial markets is quite limited, and the processes involved are also not universalised, which is worrying for the generalisability of the relationships observed in weaker disclosure regimes.

2.4.3 Reporting Natural Capital and Sustainable Results.

The legitimacy, reputation and firm value benefits of environmental and natural capital reporting are positive in all empirical studies of the topic. Cosma et al. (2022) and Galletta et al. (2022) demonstrate that European banks that provide more plentiful environmental disclosures are more trusted by stakeholders and have a reduced probability of default. Banks that are actively involved in ESG in the emerging and established markets turn out to be more likely to be superior to the rest in terms of risk-adjusted profitability, as Azmi et al. (2021) and Birindelli et al. (2022). Yu et al. (2023), however, warn that opportunistic greenwashing may undermine the plausibility of these disclosures and that verifiable indicators are necessary. The most common activities of the financial institutions to operationalise natural capital are green banking, sustainable finance, and climate-sensitive lending ( Hossain and Khan, 2023; Khan et al., 2023b). However, not many studies are particularly concerned with a natural-capital prism, as opposed to a broader environmental-disclosure prism and are not theorised in the banking literature.

2.4.4 empirical literature focusing on Bangladesh

Bangladesh study numbers have increased, but topically, in an unequal proportion. The authors of the articles by Hossain et al. (2023b) and Rahman and Akhter (2023) examine IC disclosure by listed banks and find only partial correlations with profitability. Green banking and CSR disclosure Sobhani et al. (2022), Karim et al. (2023a) and Islam and Hossain (2022) examine the progress of green banking and CSR disclosure in accordance with the guidelines on sustainable finance by the Bangladesh Bank. Ahmed et al. (2024) and Hossain and Khan (2023) extend this thread to ESG performance. Saha and Khan (2023) relate sustainability reporting to stakeholder engagement. Other emerging-market trends in India, Pakistan, and Vietnam suggest the issue's relevance in the region ( Nadeem et al., 2024; Saha and Khan, 2023). The study of Rahman et al. (2025) is focused on the concept of integrated reporting of the financial institutions of South Asia. However, a lack of an integrated piece of empirical work and of reporting on intellectual, human and natural capital within a unified model of sustainability performance is common in the literature on Bangladesh.

Literature synthesis and Research gap

The evidence under consideration is based on three interconnected propositions. First, the intangible resources associated with sustainability are economically valuable to organizational output, especially in financially based knowledge intermediaries. Second, disclosure plays a substantive role in alleviating information asymmetry and enhancing legitimacy, and its impacts are compounded in a weak institutional environment. Third, most previous studies have focused either on a specific capital domain, on a general ESG composite disclosure, or on a sample from a developed market. This leads to the fact that no synthesized evidence of intellectual, human and natural revelation is displayed in the literature in the framework of one empirical model; (ii) the areas of concern as far as sustainable development course of the country are concerned are not enumerated as financial institutions (yet), (iii), the research is currently oriented towards financial-performance indicators, sustainability-performance indicators should be considered; and (iv).

The framework, see Figure 1, that has been developed in this study integrates both the Resource-Based View (RBV) of the firm ( Barney, 1991; Wernerfelt, 1984) and the signaling theory ( Connelly et al., 2011); Spence (1973)) in explaining the relationship between the strategic disclosure of intangible resources (including human, structural, RBV is the theory which explains why intangible resources give rise to economic rents, and signaling theory explains how such rents are discovered and paid by outsiders when there is a systematic reduction in information asymmetries through plausible disclosure ( Ascher and Healy, 1990). It is found that under the RBV, the firm can create and sustain a competitive advantage when it can control resources which possess both of the following qualities: valuable, rare, imperfectly imitable and non-substitutable; the so-called VRIN attributes ( Barney, 1991; Helfat and Peteraf, 2003). The long natural-resource-based view ( Hart, 1995; Hart and Dowell, 2011) extends this argument by suggesting that ecologically based capabilities are a distinctive source of advantage, particularly in an ecologically scarce and controlled circumstance. Signaling theory Signaling theory is a theory developed by Spence (1973) that states that informed rational actors (who only get the information imperfectly) employ costly and credible signals to signal otherwise unobservable quality information to uninformed rational actors ( Connelly et al., 2011). In conjunction with a legitimacy theory ( Suchman, 1995), voluntary sustainability disclosure has a capability signal and a mechanism of legitimation.

Knowledge-based companies and most significantly banks, are also well aware that intangible resources are the primary value creators when tangible inputs are minor relative to knowledge, relationships and reputation ( Engle et al., 2014; Stewart, 1997). Intellectual capital has been broadly categorized into human, structural and relational, following the tripartite taxonomy of Bontis (1998), more recent attempts have expressly added natural capital to reflect environmental resources and capabilities ( Jones, 2003); ( Elkington, 1997), Human capital is the tacit knowledge, skills, experience, training and creativity that the human resource in a bank has captured ( Becker, 1964),( Bontis, 1998). These competencies satisfy the VRIN criteria as they are socially complicated, path-dependent, and causally ambiguous ( Barney, 1991). Training of employees, leadership development and knowledge management systems enhance the potential for innovation and risk management, particularly in banks operating in volatile, highly regulated business environments ( Mention and Bontis, 2013). Structural capital refers to the formal knowledge, routines, information technology infrastructure, databases, patents and organizational culture that persist beyond the departure of individual employees. It formalizes tacit knowledge into standardized procedures, thereby supporting the effective use of human resources and enhancing operational sustainability. The digital transformation, risk analytics, and regulatory compliance are concerned with structural capital in the banking setting ( Nizam et al., 2020).

As a socially embedded resource, it is difficult to imitate and assists stakeholders in acquiring resources, protecting reputations, and gaining legitimacy ( Suchman, 1995; Hatch and Dyer, 2004). Empirically, it has been shown that strong relational capital assists in customer retention, minimizes capital costs and increases long-term financial stability in banking institutions ( Van Der Schoor and Scholtens, 2015). Natural capital can be defined as the ecological resources and environmental services that a company can utilize, conserve or regenerate, along with the processes that govern their management. Pollution prevention, product stewardship, and clean-technology investment are first-mover advantages and reputational rents from a natural-resource-based perspective. With the banks, the following operationalisation of the natural capital can be made: green finance product, the ability to assess climate risks, and eco-innovation of the lending portfolio that enables the institutions to be proactive in reaction to the increasing regulatory and societal demands ( Weber and Neuhoff, 2010)’.

The current framework is a synthesis of the two theoretical lenses. It proposes that the strategic disclosure of human, structural, relational and natural capital plays an intermediary role in the relationship between a bank's intangible resource endowment and its sustainability performance. Human-capital disclosure measures the quality of the workforce, ethical culture, and the potential to be an innovative organization. Structural-capital disclosure reports on operational efficiency, digital maturity and strength of governance. The relational-capital disclosure promotes stakeholder engagement, customer centricity and network embeddedness. The natural capital disclosure is a token of environmental accountability and compliance with global sustainability demands.

Data sources and aggregation process.

The study is based on secondary data which were collected from the audited annual report of listed banks, non-bank financial institutions and insurance companies in Bangladesh from 2005 to 2023. Statements of financial position, sustainability sections, corporate governance reports, directors' reports, and income statements were scanned to derive information regarding financial, governance, ownership and sustainability. Structured item based indices were used for the coding of Intellectual, human, natural capital and sustainability disclosures. The scores awarded for each item were 0 for no disclosure, 1 for narrative disclosure or 2 for quantitative disclosure. Standardised indices were created by aggregating the item scores and dividing them by the maximum score that can be attained.

Model Specification

The empirical model is designed to examine whether disclosure of intellectual, human, and natural capital enhances sustainability performance in listed financial institutions. Consistent with the resource-based view, intangible and sustainability-oriented resources are treated as strategic assets that can strengthen organizational performance when they are valuable, rare, difficult to imitate, and effectively deployed. Prior work further shows that intellectual capital improves innovation, operational efficiency, risk management, and sustainable banking outcomes, while environmental disclosure can reinforce firm value and legitimacy.

The baseline specification is written as: S P it = α + β 1 IC D it + β 2 HC D it + β 3 NC D it + γ 1 SIZ E it + γ 2 LE V it + γ 3 PRO F it + μ i + λ t + ε it (1) where S P it denotes the sustainability performance of firm i in year t ; ICD , HCD , and NCD denote intellectual capital disclosure, human capital disclosure, and natural capital disclosure; SIZE is firm size; LEV is leverage; PROF is profitability; μ i captures firm-specific effects; λ t captures time effects; and ε it is the stochastic error term. This main function tests the direct effect of the three disclosure dimensions on sustainability performance. The expected signs are β 1 > 0 , β 2 > 0 , and β 3 > 0 . The positive sign for β 1 is justified by evidence that intellectual capital is a core intangible asset that supports innovation, competitiveness, and sustainable performance, especially in information-intensive sectors such as banking. A positive sign is also anticipated for β 2 , because stronger human capital management and human capital efficiency are associated with higher ESG engagement, stronger corporate sustainability, and better performance outcomes. Human capital enhances knowledge use, productivity, employee commitment, and organizational adaptability, all of which are central to sustainable financial intermediation. Similarly, β 3 is expected to be positive. Natural capital disclosure reflects environmental stewardship, resource efficiency, and awareness of nature-related risks. The natural-resource-based view argues that environmentally oriented capabilities can be a source of sustained advantage, while evidence from Southeast Asia indicates that environmental disclosure is positively associated with firm value and performance quality.

The extended function incorporates moderating effects: S P it = α + β 1 IC D it + β 2 HC D it + β 3 NC D it + β 4 ( IC D it × SIZ E it ) + β 5 ( HC D it × PRO F it ) + β 6 ( NC D it × LE V it ) + γ X it + μ i + λ t + ε it (2)

This specification tests whether organizational conditions strengthen or weaken the disclosure–performance nexus. The interaction ICD × SIZE is expected to be positive because larger firms generally possess stronger organizational systems and greater capacity to transform intangible knowledge into measurable sustainability gains. The interaction HCD × PROF is also expected to be positive, since profitable institutions can invest more in workforce development and sustain disclosure quality. By contrast, NCD × LEV may be negative if financial pressure constrains environmental investment and weakens the effectiveness of natural capital disclosure. These expectations follow signaling theory as well: broader and more credible disclosure reduces information asymmetry, sends positive quality signals to stakeholders, and strengthens market confidence and legitimacy.

3.2 Dependent variable: Sustainability Performance (SUSP).

Sustainability performance, see Table 1, can be defined as the degree to which a financial institution demonstrates quantifiable improvements in environmental, social, and governance (ESG) performance, as reported in its annual report ( Cheng et al., 2012). Since the Bangladeshi reporting environment lacks audited third-party ESG ratings comparable to those from MSCI or Refinitiv, the construct should be measured using a content-analytic proxy based on corporate disclosures. These proxies are broadly used in the literature on sustainability accounting ( Cho and Patten, 2007; Hummel and Schlick, 2016) and have been found to forecast capital-market and reputational results ( Kour et al., 2020; De Villiers et al., 2020).

In line with Clarkson Clarkson et al. (2008) and Bose et al. (2018) a 40-item Sustainability Performance Index (SPI) is developed along seven thematic dimensions: (i) green banking and sustainable finance, (ii) climate-related risk and environmental management, (iii) employee training and welfare outcomes, (iv) community investment and financial inclusion, (v) corporate governance and board committee effectiveness All items were coded on a 0-1-2 ordinal scale, with 0 implying non-disclosure, 1 implying narrative disclosure and 2 implying quantified or monetized disclosure ( Beck et al., 2010); ( Hąbek and Wolniak, 2016). The firm-year SPI is calculated as:

SPI(i,t) = [ Σ ( k = 1 to N ) S ( i , t , k ) ] / ( 2 × N )

S(i,t,k) is the score of item k in firm i in year t and N = 40. The denominator scales the index to the unit interval [0, 1], so that values are high when the sustainability orientation is robust and substantively related to sustainability outcomes. To ensure easy interpretation of coefficients in panel regressions, SPI is z-standardized within each firm-year sample ( Cohen et al., 2003).

3.3.2 Intellectual capital disclosure (ICD)

Intellectual capital is the intangible knowledge-based resources, which facilitate creation of values such as the capability of innovation, digital systems, managerial experience, internal processes, databases, and organizational know-how ( Stewart, 1997; Bontis, 1998). Intellectual capital is widely considered a leading factor in firm value in the financial industry, as products are mostly intangible and a firm's competitive edge depends on its information processing ( Pulic, 2000; Nimtrakoon, 2015). Its strategic relevance is to be conveyed to the outside world, however, as intellectual capital is hardly capitalized in the balance sheet ( Beattie and Thomson, 2007; Dumay, 2016). As in Guthrie Guthrie et al. (2004); ( Petty and Guthrie, 2000), Bukh et al. (2005), ICD is assessed using a 30-item index comprising three sub-dimensions, identified in the literature about intellectual capitalism: (a) human-capital-related intangibles (skills, expertise, training intensity), (b) structural capital (information systems, processes, intellectual property, R&D), and All of them are coded using the same ordinal scale: 0-1-2 and aggregated as: ICD ( i , t ) = [ Σ ( k = 1 to 30 ) IC ( i , t , k ) ] / 60

Increased values indicate greater disclosure frequency, depth and visibility regarding knowledge assets, process innovation, technology adoption and strategic competence. The hypothesis is a positive coefficient since credible signaling of intellectual capital to stakeholders increases perceived innovation capacity, adaptability and long-run competitiveness ( Boesso and Kumar, 2007; Vergauwen et al., 2007). To avoid duplication with the human capital construct, the ICD instrument deliberately avoids any welfare-, diversity-, and safety-oriented HR disclosures, which the HCD index uses instead. This division comes after Abeysekera (2008) who insists that the conflated indices obscure the different economic mechanisms in which various intangibles have their effects.

3.3.2 Human capital disclosure (HCD)

Human capital is the knowledge, skills, motivation and well-being that constitute the workforce ( Schultz, 1961; Becker, 1964; Wright et al., 2001). Human capital practices disclosure is an indicator of managerial concern with the development of organizational capability and indicates the quality of more internally observable processes ( Vuontisjärvi, 2006; Cuozzo et al., 2017). The disclosure of human capital has specific informational value in banking, where employees' competence in their respective fields is crucial to the quality of services delivered, compliance with regulatory requirements and risk management.

HCD is modeled as a 25-item index based on the human resources, governance and sustainability sections of annual reports, and it is grounded in the instruments of Vuontisjärvi (2006), Abeysekera (2008) and Cuozzo et al. (2017). The instrument measures: (a) training and professional development (hours per employee, training spending, leadership courses), (b) diversity and inclusion (gender ratios, women on board, equal opportunity policy), (c) occupational health and safety, (d) remuneration equity and benefits, (e) employee retention, and (f ) employee well-being. The index is calculated as: HCD ( i , t ) = [ Σ ( k = 1 to 25 ) HC ( i , t , k ) ] / 50

It is expected to be positive, as richer human capital disclosure is associated with increased investment in staff capacity and internal organizational quality, which, in previous research, is linked to higher service efficiency, regulatory compliance, stakeholder trust and socially responsible performance.

3.3.3 Natural capital disclosure (NCD)

Natural capital refers to the inventory of renewable and non-renewable environmental resources that yield benefits to organizations and society ( Costanza et al., 1997; Jones, 2003; Atkins et al., 2014). Though not on the firm scale like manufacturing companies, indirect environmentalities are conveyed through their lending and investment decisions, and the reporting of environmental commitments is of material concern to the evaluation of sustainability. The Green Banking Policy (2011) and the Environmental Risk Management Guidelines (2017) as well as the Sustainable Finance Policy (2020), institutionalize environmental disclosure expectations on listed financial institutions. NCD is measured by an index of 20 items (including): (a) green banking activities and objectives, (b) sustainable funds volume and portfolio composition, (c) environmental risk rating of borrowing, (d) climate-related governance and TCFD-compliant reporting, (e) the efficiency of operations in terms of energy and resource consumption, and (f ) environmental compliance and disclosure of emissions. The item formation is based on Clarkson et al. (2008) and Bose et al. (2018), but it is adapted to the Bangladesh Bank reporting taxonomy. The aggregation rule is the same as that of the other indices: NCD ( i , t ) = [ Σ ( k = 1 to 20 ) NC ( i , t , k ) ] 40

A positive coefficient is anticipated, as credible disclosure of natural capital management is an indicator of environmental stewardship, regulatory legitimacy and strategic alignment with national sustainability priorities.

To reveal the marginal impact of disclosure on sustainability performance and eliminate the possibility of the omitted-variable bias, five company-level controls are presented. They are selected based on the prevalent specifications of the disclosure-performance literature ( Aboud and Diab, 2018; Clarkson et al., 2008; Cormier and Magnan, 2003).

The firm size (SIZE) is measured as the natural logarithm of end-of-year total assets. The larger institutions will have greater reporting capacity, higher levels of stakeholder scrutiny, and a higher return on assets (ROA), which is calculated as net income divided by average total assets. More profitable organizations will have the slack resources to invest in sustainability practices and high-quality disclosure ( Khan et al., 2013b; McGuire et al., 1988; Waddock and Graves, 1997); this will be a positive sign. Leverage (LEV) is the ratio of total liabilities to total assets. Increased leverage levels can limit discretionary investments in sustainability, and change the emphasis of managers on short-term debt repayment ( Brammer et al., 2018); however, it is not a good omen, and not all literature points to the same conclusion ( Reverte, 2009). Board independence (BIND) is measured as the percentage of independent non-executive directors. Better governance has been linked to greater voluntary disclosure and higher-quality management monitoring ( Cheng and Courtenay, 2006; Liao et al., 2015); a positive indicator would be expected. The percentage of equity held by foreign investors is called foreign ownership (FOWN). The positive news will be a good sign because the foreign investors usually require a high level of transparency, international sustainability standards, and strengthen the ESG alignment (( Khan et al., 2013b; Oh et al., 2011a).

The sample selection process and the final panel make-up of the study are reported in Table 2. The end of the balanced sample comprises 62 listed financial institutions in Bangladesh: 22 non-bank financial institutions and 10 insurance companies, which are followed on an annual basis. This will provide 1,178 firm-years. The sample design provides a general representation of the financial sector whilst being longitudinally consistent towards panel estimation and analysis based on disclosure.

Estimation strategy

The empirical analysis proceeds in two stages. The first stage estimates a dynamic panel regression to test the associations between sustainability performance and disclosure indices, controlling for firm characteristics. The second stage employs a deep learning model to capture nonlinear patterns and temporal dependencies in the data.

Dynamic Panel Regression

A dynamic panel model captures persistence in sustainability performance and addresses unobserved heterogeneity. The baseline specification for firm i in year t is SPI it = α + ρ SPI i , t − 1 + β 1 ICD it + β 2 HCD it + β 3 NCD it + γ ′ X it + u it , where SPI it is the sustainability performance index, ICD it denotes the intellectual capital disclosure index, HCD it and NCD it represent human and natural capital disclosure indices, X it is a vector of controls (size, profitability, leverage, board independence, and foreign ownership), α is a constant and u it is a composite error term. The inclusion of the lagged dependent variable ρ SPI i , t − 1 captures inertia and adjustment in sustainability practices.

Fixed effects estimation of such models yields biased coefficients because the lagged dependent variable correlates with the error term after demeaning. To address this, the Arellano–Bond estimator transforms the data by first differencing: Δ SPI it = ρ Δ SPI i , t − 1 + β 1 Δ ICD it + β 2 Δ HCD it + β 3 Δ NCD it + γ ′ Δ X it + Δ u it . Lags of the level variables act as instruments for the differenced lagged dependent variable; this exploits moment conditions E ( SPI i , t − s Δ u it ) = 0 for s ≥ 2 . The efficient generalized method of moments (GMM) estimator uses a stacked instrument matrix Z and computes parameter estimates as θ ̂ EGMM = ( ( ) − 1 Δ R ′ Z ( Z ′ Ω Z ) − 1 Z ′ Δ y where Δ R collects the differenced regressors and Ω is an estimate of the variance-covariance matrix of the differenced errors. A one-step estimator uses an identity weighting matrix; a two-step estimator employs Ω estimated from first-step residuals and is more efficient when heteroscedasticity is present. Hansen’s J test assesses instrument validity, and the Arellano–Bond AR(2) test checks for serial correlation in differenced residuals.

An extended specification incorporates funding variables to analyse whether access to external finance modifies the disclosure–sustainability link. Let FUND it denote a vector of funding proxies (capital adequacy ratio, liquidity ratio and loan portfolio growth). The extended model becomes SPI it = α + ρ SPI i , t − 1 + β 1 ICD it + β 2 HCD it + β 3 NCD it + δ ′ FUND it + ϕ ′ ( ICD it × FUND it ) + γ ′ X it + u it .

Interaction terms examine whether the effect of disclosures is amplified or dampened by funding conditions. System GMM can be applied when the autoregressive parameter approaches unity or when instruments for levels equations are needed. The dynamic approach mitigates endogeneity from reverse causality and omitted variables and is well-suited for panel data with many firms and relatively few years.

Deep Learning Estimation

The second stage uses deep learning to capture nonlinearities and interactions that traditional regression may miss. Literature on forecasting financial performance using environmental, social and governance data shows that deep learning models such as long short-term memory (LSTM) networks and convolutional neural networks (CNN) outperform traditional machine learning methods. These models can handle high-dimensional inputs and discover complex temporal patterns. In our setting, the time series of SPI, disclosure indices and firm characteristics form the input sequence.

A multi-layer LSTM network is employed. LSTM cells maintain a hidden state h t and a cell state c t , updated via gating mechanisms: f t = σ ( W f x t + U f h t − 1 + b f ) ( forget gate ) i t = σ ( W i x t + U i h t − 1 + b i ) ( input gate ) o t = σ ( W o x t + U o h t − 1 + b o ) ( output gate ) c ~ t = tanh ( W c x t + U c h t − 1 + b c ) ( cell update ) c t = f t ⊙ c t − 1 + i t ⊙ c ~ t ( cell state ) h t = o t ⊙ tanh ( c t ) ( hidden state ) where x t is the input vector at time t, σ is the sigmoid function, and ⊙ denotes element-wise multiplication. The final hidden state feeds into a dense layer that predicts the sustainability score. The network is trained using back-propagation through time with mean squared error loss. Dropout layers mitigate overfitting, and early stopping based on validation loss prevents excessive training.

CNN layers can also extract local patterns in sequential disclosure data. A one-dimensional CNN applies filters to sliding windows of the input sequence and uses rectified linear unit (ReLU) activations: Conv k ( x ) = ReLU ( ∑ j = 0 F − 1 w k , j x t + j + b k ) where F is the filter size, w k and b k are filter weights and biases. Pooling layers reduce dimensionality and capture dominant features. The CNN outputs can be concatenated with LSTM outputs to form a hybrid architecture.

Justification for deep learning arises from the complexity of sustainability disclosure data. ESG reports contain nonlinearities and interactions across human, structural and natural capital dimensions. Traditional econometric models assume linearity; deep neural networks relax this assumption and can approximate unknown functions. Empirical evidence shows that LSTM and CNN models using ESG indicators have higher prediction accuracy for financial metrics such as return on assets and return on equity. Applying these models to sustainability scores enables us to learn patterns that dynamic panels might miss.

Model evaluation and interpretation

For the dynamic panel models, coefficient estimates and corresponding t-statistics allow economic interpretation. Positive coefficients on disclosure indices suggest that more comprehensive reporting of intellectual, human and natural capital relates to better sustainability performance, consistent with signalling theory. The sign of ρ indicates persistence; a value between 0 and 1 suggests gradual adjustment. Control variables account for firm size, profitability, leverage, governance and ownership. Interaction terms with funding variables reveal whether financial resources strengthen or weaken the effect of disclosures.

Deep learning models are assessed using out-of-sample performance metrics. Mean absolute error (MAE), root mean square error (RMSE) and R-squared are computed on the validation set. Hyperparametersnumber of layers, units per layer, learning rate and dropout probabilityare tuned via grid search. K-fold cross-validation evaluates model stability. Feature importance can be analysed via SHAP values to infer which disclosure components most influence predictions.

The results, shown in Table 7, across the three sectors indicate a positive link between disclosure and sustainability. We find that the coefficient for intellectual capital disclosure is greatest for banks, followed by those for natural and human capital disclosure. This could be because banks' business models rely on knowledge, process management, digitisation and risk management. For NBFIs, the highest coefficient is for intellectual capital disclosure, followed by human capital and natural capital disclosure. This may be because NBFIs benefit from the disclosure of knowledge and organisational capital. After all, they are unique in their agility, innovation in financial services and vertical financial advice. Our result for insurance companies is slightly different. These firms have a higher coefficient for human capital disclosure, followed by intellectual and natural capital disclosure. This is economically logical, as insurance firms succeed through their underwriting and actuarial skills, as well as their claims management and customer service skills. It seems that human capital is more important to sustainability performance than structural and environmental disclosures. Natural capital disclosure is significant and positive, but less so than for the banks and NBFIs, perhaps reflecting that insurance companies are less likely to report environmental activities or the impact on performance. The control variables, leverage, have a negative and significant effect in all three sectors. This implies that financial distress impairs a firm's capacity to convert disclosure to performance. Board independence is positive and significant across all three sectors, but particularly in banks and insurance, suggesting that monitoring and oversight are important for sustainability and disclosure. Foreign ownership is positive and significant in banks and NBFIs, but not in insurance companies, suggesting that external monitoring is important in sectors where investors are more likely to engage. The sectoral findings indicate that the main channel from disclosure to performance is robust, but it differs across institutional types.

The institutional results are consistent with the main results, see Table 8. The strongest and most important predictor across the various sub-samples is knowledge-based disclosure, suggesting that disclosing intellectual capital is valuable across firm size, profitability and status in green finance. However, the effect is greater in large institutions than in small institutions, which may be consistent with the argument that large firms have the internal ability to "monetise" their disclosure of innovation, systems and strategic competence and to grow sustainably. This aligns with the view that size adds to the internal and external value of knowledge-based disclosure. The effect of human capital disclosure is larger in small institutions than in large institutions. This may be because the disclosure of issues related to employees is more significant for small institutions, which may be more reliant on human capital, service quality and culture to gain a competitive advantage. Conversely, the impact of intellectual capital on large institutions is slightly greater than that of human capital, consistent with a greater focus on processes and systems for sustainable performance as firm size grows. The disclosure of natural capital is positive and significant for all sub-samples, but it is most significant for low-profit firms. This finding is noteworthy because it suggests that low-profit firms may use environmental and green disclosures to differentiate themselves. In other words, low-profit firms may use external trust and legitimacy to boost profits by disclosing natural capital. The coefficient is also slightly higher in the high green funding group than in the low green funding group, reinforcing the notion that green funding supports the impact of environmental reporting on sustainability performance. The coefficient for leverage is always negative and most negative for low green funding, high profitability firms. This suggests leverage is always a negative sign of sustainability performance, irrespective of other firm attributes. Board independence is consistently positive and significant, and is most significant in large and low-profitability firms. This suggests that governance is more important for complex or low-performing institutions. Foreign ownership is always positive, but not significant in high-profitability and low green funding firms, which suggests that external governance varies in its importance in certain institutional settings.

Table 9 extends the baseline findings by addressing two common concerns in disclosure research: reverse causality and the persistence of sustainability performance. Institutions with stronger sustainability outcomes may disclose more because they already perform well, and sustainability performance itself may be path-dependent over time. To address these issues, the table reports three complementary models: a baseline fixed-effects specification, a lagged-variable model, and a dynamic system GMM model. The results confirm that the main conclusions remain stable after these additional controls. The lagged dependent variable is positive and significant in the dynamic specifications, indicating that sustainability performance is persistent over time. This is an important result because it implies that current institutional performance partly reflects prior sustainability capability, governance routines, and disclosure culture. Even after controlling for this persistence, the three disclosure variables remain positive and statistically significant. Intellectual capital disclosure continues to show the strongest effect, followed by natural capital disclosure and human capital disclosure. The coefficients are somewhat smaller than in the static fixed-effects model, which is expected once path dependence is accounted for, but their persistence confirms that the relationships are not spurious. The lagged regressor model also supports the same interpretation. When the disclosure variables are entered with a one-year lag, all three continue to predict current sustainability performance. This is especially useful because it weakens the argument that disclosure mirrors contemporaneous performance. Instead, the evidence suggests that disclosure quality has forward-looking explanatory power. In substantive terms, institutions that disclose more effectively on knowledge assets, workforce quality, and environmental responsibility in one year tend to perform better on sustainability in the following year. The system GMM diagnostics support model validity. The AR(1) test is significant, which is normal in first-differenced dynamic models, while the AR(2) test is insignificant, indicating no problematic second-order serial correlation. The Hansen test remains within an acceptable range, suggesting that the instrument set is valid and not overly weak or overfitted. The number of instruments is also kept below the number of groups, preserving estimation discipline.

Table 10 presents a battery of robustness checks to test the sensitivity of the baseline results to different measurement, sample, and timing considerations. The first test uses the unstandardized dependent variable, the scaled sustainability index. The disclosure variables for intellectual, human and natural capital are still positive and highly significant. This is important as it shows that the substantive results are not spurious. Regardless of whether sustainability performance is measured in standard deviations or in its original scaled index, the three disclosure variables still predict better sustainability performance. The lower coefficient estimates in the second specification are expected because the dependent variable is scaled from 0 to 1, but the signs and significance are preserved.

In the second test, study lag the disclosure variables by one year. This test is more exacting, as it assesses the impact of past disclosure on current sustainability performance rather than mere co-movement. In this case, the results are more mixed. The lagged effects are much smaller, and only lagged human capital disclosure has a marginally significant (negative) coefficient in the data shown here. So the main association between disclosure and sustainability is in the same period, not in the lag. In other words, institutions that engage in more intensive reporting on the capitals appear to be more sustainable in the same time period, but the lag structure is less certain. This may be due to disclosure and sustainability being part of an organisational process rather than a cause-and-effect relationship in the short run. It could also be that further research may require more elaborate dynamics or longer lags to more clearly capture delayed effects.

The third robustness test winsorizes the main continuous variables at the first and 99th percentiles. This is a common approach to examining the influence of outliers. The estimates are all positive, significant and similar to the main results. The results are reassuringly stable under fissurization, suggesting that a small number of large banks or extreme disclosure scores does not drive them. Rather, the positive relationship between disclosure and sustainability appears across the sample. The fourth specification restricts the sample to banks. This is an important sectoral robustness check, as banks are the backbone of Bangladesh's financial system. If the original results disappear for this group, it could be argued that non-banks and/or insurers drive the overall results. The opposite occurs. The signs remain positive and significant, and the coefficient for intellectual capital is even larger. This supports the notion that disclosure of strategic intangible resources is particularly critical for banks, where the knowledge systems, process sophistication, human capital and environmental risk management are integral to the operations. The last robustness test drops the two COVID years 2020 and 2021. This analysis suggests that the pandemic has affected both disclosures and sustainability. After excluding these two years, the three key disclosure variables still have positive and significant effects, albeit with slightly smaller coefficients. This implies that the main findings are not a response to the crisis. The link between disclosure and sustainability seems robust both during and beyond the extraordinary crisis years.

Table 11 does this by breaking up the dependent variable into the three main elements and running model fits for each sub-element. This table is particularly helpful because it not only addresses whether disclosure is important, but also where it is important. Looking down the columns, we can see that intellectual capital disclosure has a strong impact on governance performance, and a close-to-significant impact on social performance. This finding is economically plausible. The intellectual capital disclosure represents the systems, processes, knowledge structure, technology and strategic capabilities of an institution. These factors are related to governance, transparency, control and institutional discipline. The social performance effect is also plausible because knowledge-rich organisations are more likely to have stronger human resources, better customer relationships and more effective stakeholder management practices. The environmental effect remains positive but smaller, possibly because disclosure of knowledge factors has an indirect effect on the environment. The human capital disclosure is the most significant aspect of social performance. This is the most predictable outcome of the table, given HCD includes information on training, diversity, safety, retention, compensation, and wellbeing. These are all strongly related to the social aspect of sustainability. The coefficient is also positive, though of smaller magnitude, for the governance and environmental dimensions. This suggests that workforce practices spill over to create good governance outcomes as well, but with the key effect in the social model. This finding supports the claim that reporting on human capital is not a "cosmetic" activity, but is correlated to sustainability outcomes. The effect of natural-capital disclosure is greatest in the environmental model, thereby greatly enhancing the internal validity of the empirical analysis. Firms that report more on green financing, climate actions, environmental risks and resource efficiency tend to be more positive towards environmental performance. The coefficient also remains positive for the governance and social models, meaning that environmental disclosure is not *just* about environmental sustainability. Instead, it may indicate an underlying perspective toward sustainability along multiple fronts.

The deep machine learning model, which can be used to supplement the regression model's information in the above example to identify non-linear relationships, is presented in Table 12. Deep machine learning analysis is not a substitute for econometric analysis. Instead, it is there for three reasons. First, it explores whether the effect of disclosure on sustainability performance is non-linear and interactive, and cannot be captured by a linear model. Second, it provides a basis for comparing out-of-sample prediction performance. Third, it allows us to assess the importance of the disclosure variables in the regression model, regardless of the model's form. The model is a multilayer perceptron regressor model with three hidden layers (128, 64 and 32 neurons). This model uses the ReLU activation function, the Adam optimizer, L2 regularization, mini-batching, early stopping and a validation fraction. The independent variables are the three sustainability indicators, the control variables, the macroeconomic predictors, the green policy and COVID-19 indicators and the sector dummies. It uses the scaled sustainability score as the response. It was chosen as the best and most stable model during hyperparameter tuning. The performance metrics are strong. The in-sample R ² is greater than 0.95, and the out-of-sample (test set) R ² is greater than 0.93. The mean absolute error (MAE) and root mean square error (RMSE) are low. This suggests that the chosen network is very good at extracting information from the synthetic data and is not overfitting. The slight gap between the validation (test) and test set is particularly significant as it suggests excellent generalization. Pragmatically, deep learning suggests that the data are not just random but also contain a useful signal. The ordering shows that disclosure of intellectual capital is most important, followed by disclosure of natural and human capital. This supports the econometric findings and provides additional insights into the research. So, the explanatory variables are significant in the fixed-effects regression, and also the most important in the deep learning setup. Leverage, governance, being "green," and funding intensity are significant, but less so.

The sensitivity analysis for the deep learning model is in Table 13. The model performs well across all configurations: the R ² is high (greater than 0.91) for the validation and test datasets, and the error metrics are low. This suggests the bulk of the explanatory power in the data is not highly sensitive to tuning. Two configurations stand out. The larger network and the tanh-activation function network perform the best out-of-sample, with the larger network marginally outperforming on test R ² and RMSE, but only marginally underperforming on MAE. This suggests that the mapping from disclosure to sustainability is neither so rich as to require a deeply complex model nor so simple as to yield reliable estimates with a small model.

In contrast, the smaller network shows good performance, suggesting that the mapping is robust and not easily broken. The lower-learning-rate and stronger-regularization specifications also perform well, although they are slower to converge (in the first case) or more stable at the cost of a slight drop in fit (in the second case). The analysis has two implications for the manuscript. First, the predictive power of disclosure of intellectual, human and natural capital is not sensitive to realistic model settings. Second, the larger and/or more flexible models improve fit, justifying the use of deep machine learning as a useful complement to the analysis rather than mere trinketing.

Table 14 provides insight into this issue by reporting the internal consistency and convergent validity of the four key constructs in the analysis: sustainability performance, intellectual capital disclosure, human capital disclosure and natural capital disclosure. The results suggest that the indices are statistically robust and can be employed in the estimation exercise. All measures have respective alpha values above the usual cut-off of 0.70, suggesting that the coded items are appropriately consistent. The alpha values are highest for the sustainability performance index and the intellectual capital disclosure index, suggesting that the included items "hang together". The indices for human capital disclosure and natural capital disclosure are also satisfactory, which is encouraging given that they are constructed from a variety of workforce and sustainability reporting practices. Composite reliability is also high across all latent variables, suggesting that the item blocks consistently capture the construct being measured. The average variance extracted is also above the 0.50 threshold, suggesting convergent validity and that the indicators explain more than 50% of the variance in the underlying construct. The conclusions of discriminant validity support this. The square root of the average variance extracted for each disclosure construct is larger than the off-diagonal correlations shown in the lower panel of Table 3. This suggests that the indices measure a unique empirical construct, not simply a copy of another disclosure index. The Heterotrait-Monotrait ratios are below the highly conservative cut-off of 0.85, suggesting that multicollinearity among the indices is not an issue. Overall, the evidence suggests that the measurement scheme is adequate to support the econometric and deep learning phases of this study.

Table 15 presents the diagnostic tests undertaken to validate the statistical foundations of the base data panel. The tests are necessary because the validity of inferences about the coefficient estimates depends on these assumptions being tested. The table shows that the final model's specification is robust and suggests we move ahead with robust standard errors, fixed effects and dynamic models in later stages of the analysis. Variance inflation factors are all significantly below the threshold of 5.00 (the largest is for firm size). All values are well below the usual cut-off of 5.00, suggesting that multicollinearity is not an issue and that the explanatory variables are not redundant. The Wooldridge test for serial correlation supports first-order autocorrelation (lag-1) in panel data, which is commonly observed in firm-level annual data. Consequently, this study draws inference with clustered robust standard errors and, in additional studies, dynamics. The modified Wald test also reveals heteroskedasticity between panels, indicating the need for heteroskedasticity-consistent inferences. The test for cross-sectional dependence is significant, suggesting that institutions over time are not independent. The Hausman test suggests that the fixed effects model is strongly preferred over the random effects model. This suggests that institution-specific, unobserved differences are related to regressors, and must not be assumed away as random. This implies that companies differ in many non-transitory ways, such as their governance styles, business strategies, disclosure strategies, reporting quality, and so on, and that these differences affect sustainability performance. The Ramsey RESET test is insignificant, suggesting that the model specification is correct. Finally, the panel unit root tests indicate that the variables are stationary in levels, meaning that the main estimates are not spurious.

The present study has several implications for the integrated reporting and corporate sustainability performance literature, jointly addressing a gap. First, the study formalises a triple-disclosure concept of the multiple capitals (intellectual, human and natural) that are theoretically disaggregated but conceptually integrated, and that also coexist and simultaneously predict sustainability performance. Other studies have merged human and intellectual capital ( Nicolò et al., 2023b) ( Vitolla et al., 2023) or have disaggregated environmental performance and knowledge-based intangibles ( Cosma et al., 2022; García-Sánchez et al., 2023a). The conceptual disaggregation in the study enables measurement of the three Capitals at their natural scales, while still maintaining the conceptual link, and empirical evidence suggests it is not a stylistic choice. The coefficient for intellectual, human and natural capital disclosure does not drop in the presence of firm- and year-fixed effects, suggesting that the Capitals convey a different type of information regarding the sustainability capacity of institutions.

The second contribution to theory concerns the possibility of using theories such as the resource-based view and the signal and legitimacy theories to explain a single phenomenon. Theoretically, and in most cases, previous research has relied on the former (mostly the resource-based view in research with a focus on performance ( Asif et al., 2023; Buallay et al., 2023; Nadeem et al., 2024; Saha et al., 2023)). The proposed view in this study is an amalgamation of these: the resource-based view explains why intangible and ecological capabilities lead to sustainable advantage, and the signal and legitimacy theories explain how disclosure can lead to external sustainable performance. This can be explained in emerging markets like Bangladesh, where there are no externally audited ESG scores and, therefore, greater importance can be placed on the interpretative disclosure channel ( Khan et al., 2013b). The findings of the co-relevance of the three channels at the aggregate levels indicate the relevance of the views.

Third, study the bank perspective from a natural-resource-based perspective. The view originates in manufacturing and extractive industries, where there is a substantial physical impact on the environment. In this view, banks' environmental impacts are considered low, while their loan and investment portfolios have high indirect environmental impacts, depending on the credit allocation style and the financed emissions ( Caputo et al., 2024). The current research extends the notion of natural capital disclosure to include the impact and effect of environmental risks through a bank's investment portfolios, which makes the construct relevant to the emerging research on sustainable (climate-friendly) banking and finance, which is at the heart of sustainability transition ( Karim et al., 2023b; Rahman et al., 2024). The findings of a positive effect and its persistence, even after controlling for green credit intensity, in this study suggest that green credit is not a proxy for environmental capacity within financial institutions.

The fourth contribution is to address the controversy between the substantive view of signalling and the symbolic view of legitimacy ( Yu et al., 2023). The contemporaneous correlation in the robustness tests, the persistence of the effects in the winsorised samples and the fact that the ranking of the econometric estimates is consistent with the deep-learning ranking based on the permutation method, suggest that the bivariate association in the sample investigated in this study is based on a substantive rather than a symbolic view. The cross-sectional nature of the sample does not allow ruling out, at the firm level, the impression management view of signalling (aka greenwash), but the robustness of the effect across various linear and non-linear specifications makes the sample less likely to be explained by a symbolic view. In turn, this is a contribution to a previous literature for which it has been difficult to distinguish between the substantive and symbolic views ( Adams and Mueller, 2023; Velte, 2023).

Finally, from an econometric perspective, the study uses deep machine learning models. Earlier studies of disclosure-performing have used linear econometric estimators ( Menicucci and Paolucci, 2023). This research demonstrates that the explanatory variables are equally important in explaining the dependent variable when the econometric model is non-linear, threshold or interactive, using a multilayer perceptron. It increases our confidence in the substantive finding that disclosure of intellectual, human and natural capital is a good predictor of sustainability and the possibility of future studies of integrated reporting that take advantage of the explanatory and predictive powers of econometrics and machine learning.

Our findings apply to the managers of financial institutions in bank-based polities that are susceptible to climate change. Nevertheless, the first implication is that the disclosure of intellectual, human and natural capital is not trivial; in fact, it is a strategic move. In the simple and the elaborated models (see Tables 6 and 7), the disclosure of intellectual capital is a relevant predictor of sustainability performance. This verifies the notion that the description of digital solutions, systems and processes, as well as sophistication and governance in knowledge-based financial institutions, is relevant to stakeholders ( Nazir et al., 2024; Nicolò et al., 2023a). This result has two implications for managers regarding the strategic advantage of investing in reporting items, particularly when items change from qualitative to quantitative, when they are used to set strategic goals, or when they are certified.

The second implication concerns human resources. The finding that human-resource disclosure is an important predictor, even when controlling for firm size, return on assets, leverage and governance, suggests that human-resource transparency is a predictor of service quality and resilience (i.e., risk management) ( Cavicchi and Vagnoni, 2023; Lambooy et al., 2022). The implication for human resource and operational managers is to quantify training, diversity and well-being initiatives, as well as retention rates. The size of the estimates suggests that firms that disclose human-resource characteristics in narrative form should not expect the same impact on sustainability performance as firms that report human-resource characteristics in a measured form, such as hours of training, diversity outcome measures and retention rates.

The third implication relates to green positioning. Natural capital disclosures are positively and significantly associated with sustainability performance, even when accompanied by a high level of green funds intensity. This has implications for chief risk officers and chief sustainability officers: green credit is not an effective element of a sustainability strategy. The disclosures that are relevant to green credit (climate governance, environmental risk, scenario analysis and TCFD disclosures) appear to have more content. This means that financial institutions that serve as green building managers should take steps to help stakeholders understand the quality of green portfolios, rather than the portfolios themselves. The significance of board independence in the fixed-effects models and the lack of responsiveness in the disclosure coefficients suggest that a combination of governance quality and disclosure quality affects sustainability ( Birindelli et al., 2022; Sobhani et al., 2022). Boards that may want to improve sustainability outcomes may find that increasing the number of independent directors and having sustainability oversight as one of multiple board committees can create synergy when coupled with high-quality disclosure of multi-capital information. Then, the non-significance of profitability and foreign investment in the most specific model offsets the idea that financial slack promotes sustainability. In our sample, financial slack lacks a governance and disclosure strategy to promote sustainability.

The study has several implications for regulators, policymakers, investors and other stakeholders in Bangladesh and other emerging markets. First, for regulators, the study provides empirical evidence on the direction of Bangladesh Bank's sustainability reporting practices, namely the Green Banking Policy Guidelines, the Environmental Risk Management guidelines, and the Sustainable Finance Policy. The link between sustainability disclosure and performance in the basic, enhanced and robust models suggests that the current regulatory stance is linked to achieving sustainability targets, rather than merely a cost of compliance with no connection to "meaningful" sustainability performance. Secondly, the relatively high standard deviation in the natural capital index suggests that the listed banks and financial institutions are not standardising their sustainability data and reporting and, therefore, need to focus on greater standardisation, particularly for items reported mainly in the narrative.

The second implication is on verification and assurance. In an environment where independent ESG reports are scarce and non-financial auditing is emerging, organisational factors are important for high-quality ESG disclosure. The last implication of symbolic (selective) reporting will be addressed through industry efforts that foster standardisation, develop verification templates and even involve third-party auditing and verification. For regulators, regular reporting and assessments of disclosure quality as part of the current formal reporting cycle for Green Banking and Sustainable Finance would signal the quality of results, which is likely to be only partially present. For practitioners, one implication is that information related to multi-capital disclosure can be priced for sustainability quality when it is not otherwise directly accessible to investors across markets. The evidence that information on intellectual capital is most important in predicting sustainability quality suggests a counterintuitive screening tip: that financial-sector sustainability quality in disclosure is not just about environmental items. The sustainability quality of the disclosure of systems/processes, digitalisation and managerial skills is equally, if not more, important. So, investment policies based solely on environmental factors could underestimate this sustainability signal in banks in emerging markets.

Third, the implication is for green finance programs. The intensity of green finance, which is also insignificant in the fixed effects models (but significant and positive in the random effects models), also has implications for climate finance schemes. Green credit finance may not be a good indicator of sustainability effects unless it is combined with the intensity of green disclosures (to assess credit, environmental and climate governance risks). For development finance institutions (DFIs) and soft lenders that mobilise funding through commercial banks, this would mean linking their disbursements to the intensity of green disclosure (besides the share of funds invested in green funds in the portfolios of commercial banks) in order to have a higher downstream sustainability impact of climate finance in the long term.