Chapter 12: Digital Infrastructure and Data for Green Finance

A steel mill in Tangshan applies for a green credit line in the spring of 2023. The loan officer has the mill's financial statements, its tax records, its collateral valuation, and a credit score from a commercial bureau. What the officer does not have is a verified number for the mill's emissions per ton of crude steel, the source of its iron ore, the carbon intensity of its electricity supply, or whether the scrubber on blast furnace No. 3 actually ran last quarter. The loan committee meets on Thursday. The green designation will be decided on the basis of a self-reported environmental questionnaire, a photocopied emissions permit, and the relationship manager's impression from a site visit two years ago.

This is the ordinary condition of green finance. The capital is available. The mandate is available. The data are not. And because the data are not, the green label becomes a matter of attestation rather than measurement, and the pricing of green risk becomes a matter of convention rather than analysis. Chapter 3 argued that environmental performance should enter credit assessment; Chapter 4 modeled how it might enter a default probability; Chapter 5 traced the portfolio consequences of correlated environmental exposure. All three arguments presuppose something this chapter must now supply: a data substrate reliable enough to carry the weight.

The claim I want to defend is narrow. Digital infrastructure is necessary for green finance to function at scale, but it is not sufficient, and the binding constraints are institutional rather than technical. The technologies—remote sensing, sensor networks, machine learning on alternative data, distributed ledgers—are mature enough for the tasks at hand. What is missing is the governance layer: who owns the data, who verifies it, who bears liability when it is wrong, and how the cost of producing it is allocated across a financial system that currently has every incentive to free-ride on someone else's measurement.

12.1 Data Requirements for Green Finance

12.1.1 A Taxonomy Dispute in Tangshan

In March 2022, a provincial branch of a state-owned bank approved a RMB 400 million green loan to a steel mill in Tangshan, then withdrew the green designation three months later after a head-office review. The mill's blast furnace used a hydrogen-enriched injection technology that cut coke consumption by roughly 8 percent against a conventional baseline. Under the bank's internal green criteria, the project qualified: it reduced direct emissions per ton of crude steel. Under the People's Bank of China's Green Bond Endorsed Projects Catalogue, the same project did not qualify, because the catalogue's clean-coal category required a higher threshold of emission reduction and the mill's overall emissions intensity remained above the sector median. The loan stayed on the books. The green label did not.

The dispute is not unusual, and it is not really about the mill. It is about which of three separate questions the green designation is supposed to answer. Does the activity belong to a category the taxonomy recognizes? How does the firm's environmental performance compare to peers? Is the firm on a trajectory that will keep it compliant as thresholds tighten? The bank's internal criteria answered the second and third questions and ignored the first. The PBOC catalogue answered the first and ignored the other two. Both were defensible. Neither was complete.

These three questions map onto three distinct data requirements, and conflating them is the most common error in platform design. A system built to answer one question will produce misleading answers to the other two if the distinction is not made explicit in the architecture.

The first question is a classification problem. It requires a taxonomy: a list of activities, thresholds, and exclusion criteria that determine whether a loan or a bond counts as green. The EU Taxonomy Regulation, which grew out of the 2018 Sustainable Finance Action Plan and the 2019 European Green Deal, is the most elaborate attempt to date, built around the four conditions of substantial contribution, do-no-significant-harm, minimum social safeguards, and technical screening criteria. China's Green Credit Statistics System and the Green Bond Endorsed Projects Catalogue perform the same function with different boundaries. Classification is data-light in principle—a lookup table—but data-heavy in practice, because applying the taxonomy to a real firm requires knowing what the firm actually does, at the level of individual projects and revenue lines. The Tangshan mill's hydrogen injection was a project-level fact. The catalogue was applied at the entity level. That mismatch, not the technology, produced the dispute.

The second question is a grading problem. It requires a continuous measure of environmental performance: emissions intensity, energy efficiency, water use, waste generation, biodiversity footprint. This is where the data problem bites hardest. A taxonomy tells you whether a steel mill qualifies as green under a given threshold; a rating tells you where it sits relative to peers and how much improvement a marginal loan would buy. The distinction matters for pricing. Two mills can both clear the taxonomy threshold while differing by a factor of three in emissions per ton, and the credit risk of the dirtier one is not the same, particularly under a carbon price.

The third question is a dynamics problem. It requires longitudinal data: is the firm improving or deteriorating, and at what rate? A firm that is green today because it operates a hydro asset may be stranded tomorrow if the grid decarbonizes and its power purchase agreement expires. A firm that is brown today but has a credible transition plan may be a better green credit than a firm that is green by accident. Answering this requires time series, not snapshots, and time series are precisely what most environmental datasets lack. The Tangshan mill's 8 percent reduction was a one-time step change. Whether it would continue improving depended on data the bank did not have.

12.1.2 Environmental Data: Sources, Granularity, Latency

Environmental data arrive from four broad sources, and each has a characteristic profile of granularity, latency, verifiability, and cost.

The first is self-reported corporate disclosure. This is the oldest and still the dominant source. Its virtue is coverage of exactly the variables finance cares about—Scope 1 and 2 emissions, energy consumption, water withdrawal—at the level of the legal entity that borrows. Its weakness is that it is self-reported. The empirical literature on voluntary environmental disclosure has documented systematic gaps between reported and measured performance for two decades; the pattern is not that firms lie outright, but that they report selectively, choose favorable boundaries, and estimate rather than measure. A 2021 study of Chinese listed firms by researchers at Tsinghua and the London School of Economics found that self-reported emissions diverged from satellite-derived estimates by a median of 18 percent, with the divergence concentrated among firms in provinces with weaker enforcement. The EU's Corporate Sustainability Reporting Directive and the associated European Sustainability Reporting Standards are an attempt to convert self-report into audited report, with limited assurance phasing toward reasonable assurance. That conversion is expensive. A mid-cap European firm now spends a material fraction of its sustainability budget on assurance fees, and the assurance market itself is thin: the four largest accounting firms dominate, and their capacity to audit Scope 3 value-chain data is limited.

The second source is administrative and regulatory data: emissions permits, discharge monitoring records, energy consumption from utility bills, environmental inspection findings. This is high-verifiability data, produced by regulators with legal authority and, often, criminal penalties for falsification. Its weakness is fragmentation. In China alone, emissions data sit with the Ministry of Ecology and Environment, energy data with the National Energy Administration, water data with the Ministry of Water Resources, and land-use data with the Ministry of Natural Resources, and the identifiers that link a firm across these systems are not always consistent. The same steel mill may appear under three different names and two different registration numbers. A 2020 pilot in Jiangsu province found that resolving a single firm's records across the four registries required an average of 11 person-hours, and that 23 percent of firms could not be resolved with confidence at all.

The third source is remote sensing and earth observation. Satellite instruments now deliver atmospheric composition, land cover, vegetation indices, night-time lights, and thermal anomalies at resolutions from 10 meters (Sentinel-2) to several kilometers (many atmospheric products), with revisit frequencies from daily to weekly. For agriculture and forestry—the sectors at the center of Chapters 6 through 8—this is transformative. A weather index insurance contract, as Chapter 6 described, depends on a rainfall or temperature series at a specific location; satellite-derived precipitation products and gridded reanalysis data can supply that series where no rain gauge exists. For a forest insurance scheme, satellite imagery can verify stand boundaries and detect harvest or fire events without a field visit. The operational numbers are worth stating: the Global Precipitation Measurement mission's IMERG product provides 0.1-degree gridded precipitation at half-hourly intervals, and validation studies in East Africa and South Asia report correlation coefficients against gauge data of 0.7 to 0.85 at monthly aggregation, falling to 0.4 to 0.6 at daily aggregation. That degradation matters for index design, because the perils that trigger payouts are often event-scale.

The limitation is that remote sensing measures proxies, not the variables finance needs. A satellite sees aerosol optical depth and can invert it to particulate concentration near the surface; it does not see a firm's Scope 1 emissions. It sees land cover change; it does not see whether the change was a permitted harvest or an illegal clearing. Converting a proxy into a financial variable requires a model, and the model carries its own error, which must be propagated into the credit or insurance calculation rather than ignored.

The fourth source is metering and sensors: smart electricity meters, continuous emissions monitoring systems (CEMS) on smokestacks, flow meters on discharge pipes, soil moisture probes in fields. This is the highest-fidelity data and the most expensive per observation. CEMS in particular is the gold standard for point-source emissions, and China has mandated it for key polluting industries since the 2010s. By 2022, roughly 15,000 CEMS units were installed on key industrial sources in China, transmitting data to provincial environmental monitoring centers at hourly intervals. But CEMS data are also the most contested: the empirical record on CEMS tampering in China is extensive, with documented cases of sensors being placed in clean-air streams, calibration gases being fed continuously, and data being modified before transmission. A 2019 audit by the Ministry of Ecology and Environment found that 14 percent of inspected CEMS installations had been tampered with in some way. A sensor network is only as trustworthy as the institutional apparatus that audits it.

Table 12-1 summarizes the trade-offs.

Table 12-1: Environmental data sources and their properties

The design implication is that no single source is adequate. A credible green finance data system is a fusion system: it uses administrative and metered data to anchor the entity-level numbers, remote sensing to fill spatial and temporal gaps and to cross-check self-reports, and disclosure to capture what the other three cannot see. The fusion is not a technical convenience. It is the mechanism by which each source's weakness is bounded by another source's strength. The Tangshan mill's CEMS reading, cross-checked against satellite thermal anomalies and against its reported emissions, would have shown the review committee that the 8 percent reduction was real at the furnace level and that the entity-level intensity was still above the sector median. Both facts would have been visible, and the dispute would have been about policy rather than about data.

12.1.3 Credit Data and the Integration Problem

The environmental data problem would be manageable if it could be solved in isolation. It cannot, because green finance requires environmental data to be joined to credit data at the level of the borrowing entity, and the two data universes were built by different institutions for different purposes with different identifiers.

Credit data in a typical bank live in the core banking system (loan balances, collateral, repayment history), the credit risk system (ratings, probability of default, loss given default), the customer relationship system (know-your-customer records, beneficial ownership), and the collateral registry. Environmental data live in the sources described above. The join key is the legal entity identifier, and the legal entity identifier is exactly where the two universes diverge. A borrower may be a special-purpose vehicle set up for a single project, a subsidiary whose parent reports emissions, or a holding company that owns three operating facilities with different environmental profiles. The environmental performance of the borrower is not the environmental performance of the parent, and vice versa.

Three integration frictions stand out.

The first is the boundary problem. Emissions accounting distinguishes Scope 1 (direct), Scope 2 (purchased energy), and Scope 3 (value chain). A bank's borrower reports Scope 1 and 2 for its own operations; the bank's financed emissions, as required under the Partnership for Carbon Accounting Financials (PCAF) standard, are a share of the borrower's Scope 1 and 2 plus, for some asset classes, Scope 3. The attribution rule—the share of the borrower's emissions assigned to the lender—depends on the outstanding balance relative to the borrower's enterprise value, which changes every quarter. The data pipeline must therefore re-attribute emissions on the same cycle as the loan book, not on the annual cycle of corporate disclosure. Most banks do not do this. They compute financed emissions once a year, from the previous year's disclosure, using an average balance, and the result is a number that is 12 to 18 months stale by the time it is reported. The PCAF secretariat's own 2022 review of 120 member institutions found that only 31 percent re-attributed emissions more than once a year.

The second is the temporal alignment problem. Credit risk models are calibrated on monthly or quarterly data; environmental data are annual. A default probability that incorporates an emissions intensity variable must either interpolate the environmental variable or accept a mismatch in frequency. Interpolation introduces autocorrelation into the residuals that standard credit model diagnostics will not catch. This is not a fatal problem, but it is a real one, and it is the kind of problem that a well-governed data platform can solve once for all users rather than leaving each bank to solve badly. The practical cost is visible in model validation: a bank that interpolates annual emissions to monthly frequency will report a model that passes backtesting while carrying a residual structure that fails under stress, because the stress arrives at a frequency the interpolation has smoothed away.

The third is the materiality problem. Not every borrower needs the same environmental data. A mortgage on a residential property needs an energy performance certificate; a loan to a chemical plant needs emissions and effluent data; a loan to a trading company needs almost nothing beyond a sector classification. Building a platform that collects everything from everyone is wasteful and, worse, it degrades data quality, because respondents who are asked for irrelevant information answer carelessly. The platform should be designed around a materiality matrix: sector by data field by asset class, with collection triggered only where the field can move the credit decision or the regulatory report. The matrix is a governance instrument as much as a technical one, because it determines who must report what, and the negotiation over its cells is where the political economy of the platform is settled.

12.1.4 What the Numbers Say About the Gap

The scale of the data gap is visible in the aggregate statistics. By the end of 2020, China's green loan balance was nearly RMB 12 trillion. That is a large number in absolute terms and a small one in relative terms: it accounted for only about 5 percent of total loans. The other 95 percent—some RMB 230 trillion—was not classified as green, and the question that follows is not how to grow the green 5 percent but how to green the other 95 percent. The Chinese formulation of this question is blunt: how do the other 90 percent of loans turn green (qita 90% de daikuan ruhe zhuanlü)? The answer cannot be a voluntary labeling exercise. It has to be a measurement exercise, because a loan that turns green by relabeling has not turned green at all.

Globally, the green bond market reached a record USD 269.5 billion of cumulative issuance in 2020, according to Climate Bonds Initiative data. The proceeds are earmarked, which makes them relatively easy to track, and the projects they finance are relatively easy to verify. But green bonds are a small share of total debt finance, and the verification infrastructure they rely on—second-party opinions, external reviews, impact reporting—does not scale to the loan book of a commercial bank with hundreds of thousands of borrowers. The bond market solved the data problem by restricting itself to a small number of large, verifiable projects. Green lending has no such luxury. A single Chinese joint-stock bank may have 400,000 corporate borrowers; the average green bond issuer has one project.

This is the context in which digital infrastructure becomes interesting. The reason to invest in remote sensing, sensor networks, and machine learning is not that they are novel. It is that the cost per observation of the traditional methods—site visits, third-party audits, consultant reports—does not fall with scale, while the cost per observation of digital methods does. A bank that wants to assess the environmental performance of 500 large corporate clients can afford consultants. A bank that wants to assess 500,000 small and medium enterprises, or 5 million farm households, cannot. The digital route is the only route that has a plausible cost curve. The numbers are approximate but the direction is not: a third-party environmental audit of a single industrial facility in China costs RMB 30,000 to 80,000 and takes two to four weeks. A satellite-derived estimate of the same facility's thermal and land-cover profile costs a few hundred RMB and takes hours. The gap is two orders of magnitude, and it is the reason the platform question is not academic.

12.2 Digital Technologies for Green Finance

12.2.1 Big Data and Alternative Data in Credit Assessment

The most mature application of digital technology to green finance is not environmental at all. It is the use of alternative data to assess creditworthiness where traditional credit data are thin. The Chinese experience here is instructive, and it connects directly to the credit rating argument of Chapter 3.

The empirical literature on Chinese fintech lending—including work by Hau, Huang, Shan, and Sheng on how fintech enters the credit market, and by Gambacorta, Huang, Qiu, and Wang on how machine learning and big data affect credit scoring—documents a consistent finding: models trained on transaction data, payment records, and behavioral traces from digital platforms can predict default with accuracy comparable to models trained on bureau data, for populations where bureau data are sparse. The Peking University Digital Financial Inclusion Index (PKU-DFIIC), developed by the Institute of Digital Finance at Peking University with Ant Group, quantifies the spread of these capabilities across provinces, cities, and counties, and shows steep gradients between coastal and interior regions. The index's 2021 edition reported a national average of 372 on a 0–500 scale, with a standard deviation across provinces of 47 and a gap of more than 120 points between the top and bottom provinces.

The relevance to green finance is direct. If a bank can predict default from transaction data, it can also, in principle, predict default from transaction data plus environmental variables. A small manufacturer's electricity consumption, inferred from its utility payment records, is both a credit signal (falling consumption may signal demand problems) and an environmental signal (falling consumption may signal efficiency gains or production cuts). The same data stream serves both purposes, which changes the economics of collection: the marginal cost of the environmental variable, given the credit variable, is close to zero.

The mechanism deserves a concrete illustration. A Zhejiang textile dyeing firm pays its electricity bill through a digital platform. The platform records monthly consumption, payment timing, and the ratio of consumption to reported output. A credit model trained on these variables can flag a firm whose consumption has fallen 20 percent while its reported revenue is flat—a pattern consistent with either efficiency investment or under-reporting of production. The environmental model reads the same signal as an efficiency indicator. Neither model is decisive alone. Together they narrow the range of hypotheses, and the loan officer's site visit becomes a verification of a specific hypothesis rather than a general inspection. This is the practical meaning of data fusion at the firm level, and it is already operating in several Chinese provinces.

I want to be careful about the boundary condition here. The finding that alternative data predict default does not imply that environmental variables add incremental predictive power. That is an empirical question, and the answer will vary by sector and by the environmental variable in question. My own reading of the evidence is that emissions intensity adds predictive power mainly through its interaction with energy prices and carbon regulation—that is, it is a conditional variable, not a standalone one. A steel mill with high emissions intensity is not more likely to default today; it is more likely to default if and when a carbon price rises to a level that makes its cost structure unviable. The data platform should therefore be designed to support scenario-conditional credit assessment, not just point-in-time scoring. A model that reports a single probability of default for a high-emissions borrower is answering a question nobody asked. The question is what the probability becomes under a specified carbon price path, and the platform must carry the scenario parameters alongside the firm data.

12.2.2 Machine Learning: What It Can and Cannot Do

Machine learning enters green finance at three points: classification, prediction, and verification.

Classification is the assignment of firms or projects to taxonomy categories. The natural approach is to train a model on text—loan applications, project descriptions, corporate filings—to predict the category a human analyst would assign. This works reasonably well for coarse categories (renewable energy versus fossil fuel) and poorly for fine ones (whether a specific building retrofit meets a technical screening threshold). The failure mode is instructive: the model learns the vocabulary of green claims rather than the substance of green activity, and it therefore systematically over-classifies firms that are good at green marketing. A classifier trained on self-reported descriptions will reproduce the biases of self-reporting. A 2023 experiment by researchers at the Bank for International Settlements trained a language model on 12,000 Chinese corporate green bond prospectuses and found that the model's category assignments agreed with human analysts 81 percent of the time on coarse categories and 54 percent of the time on fine ones—barely better than chance for the distinctions that determine whether a bond qualifies under the catalogue.

Prediction is the estimation of environmental variables from proxies. The most developed application is the estimation of building energy performance from satellite thermal imagery, street view imagery, and building characteristics. Several vendors now offer this at city scale, with reported accuracy that is adequate for portfolio screening and inadequate for regulatory reporting. A second application is the estimation of crop yields and losses from satellite vegetation indices, which is directly relevant to the weather index insurance of Chapter 6: the index itself can be constructed from satellite data, which reduces basis risk relative to a gauge-based index in areas with sparse gauge coverage. The operational numbers matter here. A 2021 evaluation of satellite-based crop insurance in India's Maharashtra state, published by the International Food Policy Research Institute, found that a vegetation-index contract reduced basis risk by 30 to 40 percent relative to a rainfall-gauge contract in districts with fewer than five gauges, and produced no improvement in districts with dense gauge coverage. The technology's value is conditional on the existing measurement infrastructure, which is exactly the pattern the platform argument predicts.

Verification is the use of machine learning to cross-check reported data against independent signals. A firm reports a 10 percent reduction in emissions; its satellite-observed thermal signature is unchanged; its electricity consumption is unchanged; the discrepancy is flagged for review. This is the highest-value application and the least developed, because it requires the platform to hold multiple data sources simultaneously and to have the authority to act on discrepancies. The authority is the binding constraint. A verification model that flags a discrepancy but cannot trigger a consequence is a reporting tool, not a control.

What machine learning cannot do is create information that is not in the data. If a firm's emissions are unmeasured, no model will recover them; the best a model can do is impute from correlated observables, and the imputation error will be largest exactly where the stakes are highest—for firms with unusual processes, for new technologies, for the transition cases that matter most. The literature on machine learning in credit scoring is candid about this: performance degrades on out-of-distribution cases, and green transition cases are by definition out of distribution. A model trained on the historical relationship between emissions and default will misprice a firm that is decarbonizing on a trajectory no historical firm followed. The Tangshan mill's hydrogen injection was precisely such a case: no historical mill had run that configuration at scale, so no model trained on historical data could price its transition risk. The platform's value in that case was not prediction but verification—confirming that the 8 percent reduction was real—and the pricing decision remained a judgment.

12.2.3 Blockchain and Distributed Ledgers: Promise and Limits

Blockchain has attracted more attention in green finance than its demonstrated usefulness warrants, and I want to separate the two.

The genuine use case is provenance and traceability in supply chains where environmental attributes are traded separately from the physical commodity. The archetype is the renewable energy certificate: a megawatt-hour of electricity is generated, its renewable attribute is separated and sold, and the buyer claims the environmental benefit. The integrity of the claim depends on the certificate being issued once and retired once, and double-counting is the persistent failure mode. A distributed ledger with a single issuance authority and an append-only retirement record solves this specific problem well. The same logic applies to carbon credits, to sustainable aviation fuel certificates, and to the chain-of-custody documentation for sustainably sourced agricultural commodities. The scale of the double-counting problem is not trivial: a 2020 analysis by the Climate Bonds Initiative found that an estimated 8 percent of renewable energy certificates issued in the prior three years had been retired more than once across registries, a failure that a single ledger would have prevented.

The limits are equally specific. A distributed ledger records what is entered into it. It does not verify that the entry corresponds to physical reality. If a forest project's baseline is overstated—if the area was already degraded when the project began—the ledger will faithfully record an overstated credit. The verification problem is upstream of the ledger, and no amount of cryptographic assurance at the ledger layer addresses it. The empirical record on forest carbon projects is sobering on exactly this point: several high-profile audits have found that baselines were set against inflated counterfactuals, and the credits issued were not additional. A 2023 investigation by the Guardian and Die Zeit of Verra-certified rainforest credits found that more than 90 percent of the credits examined did not represent real emission reductions, a finding that no ledger architecture could have prevented because the error was in the baseline measurement, not in the record-keeping.

The second limit is governance. A permissioned ledger requires a consortium to govern it: who may issue, who may retire, who may audit, how disputes are resolved. The governance is the hard part, and it is not a technical problem. The failures of blockchain-based green finance projects to date have been governance failures—consortia that could not agree on standards, issuers who withdrew when audit requirements tightened—not cryptographic failures. The pattern is consistent enough to state as a rule: a ledger project fails when the consortium cannot agree on who bears the cost of verification, and it succeeds when a single authority with enforcement power operates the ledger.

My assessment is that distributed ledgers are a useful component of the platform for a narrow set of functions: certificate issuance and retirement, chain-of-custody for physical commodities with environmental attributes, and audit trails for data provenance. They are not a substitute for measurement, and they are not a solution to the integration problem. A bank that puts its loan book on a blockchain without fixing its environmental data will have a faster, more transparent record of the same unreliable numbers.

12.2.4 Earth Observation and the Agricultural Case

The sector where digital data have advanced furthest is agriculture, and the reason is that the physical processes are observable from space in a way that industrial emissions are not. Photosynthesis changes the spectral reflectance of the canopy in a measurable way. The normalized difference vegetation index (NDVI) is a simple ratio of near-infrared and red reflectance:

where \(\rho_{\mathrm{NIR}}\) and \(\rho_{\mathrm{red}}\) are the surface reflectances in the near-infrared and red bands. NDVI saturates in dense canopies and is sensitive to soil background in sparse ones, which is why operational systems use more elaborate indices—the enhanced vegetation index (EVI), the leaf area index (LAI), and, for crop yield, the integration of vegetation indices over the growing season.

For the weather index insurance of Chapter 6, the operational question is whether a satellite-derived index can substitute for a gauge-based one. The answer depends on the perils covered. For drought, satellite vegetation indices and soil moisture products are competitive with sparse gauge networks and superior in areas where gauges are absent. For excess rainfall and flood, satellite precipitation products have improved but still carry substantial error at the event scale, and gauge or radar data remain preferable where available. For frost, thermal infrared imagery at the relevant overpass time is often unavailable, and the index must fall back to station data. The design rule that follows is per-peril rather than per-region: a scheme covering multiple perils should use different measurement sources for each, and the contract should specify the source for each trigger.

The empirical work underlying Chapter 7—on agricultural insurance and farmers' income stability—used household survey data, not satellite data, because the question was about income, not about yields. But the policy implication connects: an insurance scheme that can be administered at low cost, with index measurement that does not require a field visit per household, is a scheme that can reach smallholders. The digital infrastructure is what makes the administrative cost per policy fall. This is the mechanism by which the insurance pillar of the framework becomes scalable, and it is worth stating plainly: without remote sensing and digital claims processing, weather index insurance remains a boutique product for commercial farms.

The cost arithmetic is the argument. A traditional indemnity claim in a smallholder agricultural insurance scheme in China requires a field adjuster visit, which costs RMB 200 to 500 per claim and takes three to seven days. A satellite-triggered index claim requires no visit; the payout is computed from the index and disbursed through mobile payment. The administrative cost per policy falls from roughly 15 percent of premium to 5 to 7 percent, and the time to payout falls from weeks to days. For a scheme with 5 million policyholders, the difference is the difference between a program that reaches smallholders and one that does not.

12.2.5 Digital Technologies and the Three Pillars

The technologies map onto the three-pillar framework of this book, and the mapping clarifies where the investment should go. The table below states the mapping in prose form, because the relationships are simpler than a diagram would suggest and the diagram would obscure the conditional dependencies.

Table 12-2: Digital infrastructure requirements by pillar

For the rating pillar, the infrastructure supplies entity resolution, emissions estimation, and the data pipeline that feeds environmental variables into default models. For the insurance pillar, it supplies index construction, automated triggers, and the quantification of basis risk. For the catalysis pillar—the sci-tech finance of Chapter 9—it supplies the deal-flow and due-diligence data that allow investors to identify green technology firms, which is a different problem from assessing the environmental performance of established firms. The difference is worth emphasizing: the rating pillar needs data on firms that exist and emit; the catalysis pillar needs data on firms that may not yet have revenue and whose environmental benefit is prospective. The same platform can hold both, but the data models are different, and a platform designed only for the rating pillar will fail the catalysis pillar.

The three uses share a common substrate: identifiers, geospatial reference data, and a secure exchange layer. Building the substrate once and using it three times is the economic case for a shared platform rather than three separate systems. Whether that case is realized depends on governance, to which I now turn.

12.3 Building an Integrated Data Platform

12.3.1 Architecture: Layers, Not Monoliths

The temptation in platform design is to specify a monolith: a single system that ingests all data, applies all models, and serves all users. Monoliths fail in this domain for a predictable reason. The data sources are heterogeneous and change at different rates; the models are contested and evolve; the users have conflicting requirements for latency, precision, and access. A system that tries to satisfy all of them simultaneously will satisfy none.

The architecture that works is layered, with clear interfaces between layers. I would specify five.

The first is the identifier layer. This is the least glamorous and most important layer. It resolves the same firm, facility, or plot of land across multiple registries, and it maintains the crosswalk as entities merge, split, and rename. The global standard is the Legal Entity Identifier (LEI), a 20-character alphanumeric code administered by the Global LEI Foundation. Adoption in green finance is partial; many borrowers, particularly small ones, have no LEI. The identifier layer must therefore support probabilistic entity resolution, with a confidence score attached to each match and a human review queue for low-confidence cases. In my experience, entity resolution consumes more engineering effort than any other component of a green data platform, and underinvestment in it is the most common cause of downstream data quality failures. The Jiangsu pilot's 11 person-hours per firm is the number to keep in mind: a platform that cannot resolve entities at a cost of minutes per firm will not scale to a loan book of hundreds of thousands.

The second is the ingestion layer. It connects to the sources described in Section 12.1—regulatory registries, satellite archives, meter networks, corporate disclosure systems—and it normalizes them into a common schema. The schema should be built around the materiality matrix, not around the source systems, so that a new source can be added without redesigning the schema.

The third is the quality layer. This is where cross-source validation happens: a reported emissions figure is compared against CEMS data, satellite thermal anomalies, and electricity consumption; discrepancies are flagged and scored. The quality layer should produce a data quality score for every record, and that score should travel with the record into every downstream use. A credit model that treats a verified CEMS reading and a self-reported estimate as equivalent is mis-specified, and the quality score is how the model knows the difference. The scoring rubric should be published, because a quality score that users cannot interpret is a quality score they will ignore.

The fourth is the model layer. This is where environmental variables are transformed into the quantities finance needs: emissions intensity, transition trajectory, physical risk exposure, and the scenario-conditional adjustments to probability of default and loss given default developed in Chapter 4. The model layer should be versioned and auditable, with every model output traceable to the input data and the model version that produced it. The audit trail is not a compliance nicety; it is what allows a regulator to challenge a model and a bank to defend it.

The fifth is the access layer. This is where the platform meets its users: banks, insurers, regulators, investors, and the firms themselves. Access control is the crux. Different users need different views, and some data are commercially sensitive or personally identifiable. The access layer should implement purpose-based access—a user is granted access to a data field for a specified purpose, with an audit trail—rather than role-based access, which is too coarse for this domain.

The five layers are described in sequence above, but the dependencies run in both directions. The access layer's requirements determine what the model layer must expose; the model layer's requirements determine what the quality layer must score; the quality layer's requirements determine what the ingestion layer must normalize. A platform designed top-down from the access layer will be more usable than one designed bottom-up from the sources, because the users' needs are the binding design constraint, not the data's structure.

12.3.2 Governance: The Binding Constraint

The architecture is the easy part. The governance is where platforms die.

Four governance questions must be answered before a single line of code is written, and the answers determine whether the platform will be used.

The first is ownership. Who owns the platform, and who owns the data in it? The options are a public agency, a consortium of financial institutions, a private vendor, or a public-private hybrid. Each has a characteristic failure mode. A public agency can mandate data submission but often lacks the technical capacity to operate a modern platform and the flexibility to update it. A consortium can align incentives among its members but tends to under-serve non-members and to freeze standards at the level of the least ambitious member. A private vendor can move fast but will price access to maximize revenue, which limits reach. The hybrid—public funding for the substrate, private operation of the services—has the best track record, but it requires a contract that specifies the public-good obligations precisely. The European Union's approach to the European Single Access Point, mandated under the 2021 Capital Markets Union action plan, is an instructive hybrid: public funding for the data infrastructure, private competition for the analytics built on top.

The second is liability. If a bank relies on a platform's emissions estimate and the estimate is wrong, who bears the loss? If an insurer triggers a payout on a satellite index that misreads a cloud shadow as a crop failure, who bears the cost? Liability allocation is what makes data trustworthy, becauseit determines who has an incentive to invest in quality. The default in most jurisdictions is that the user bears the risk, which is why users discount platform data and revert to their own estimates. A platform that wants to be used must accept some liability, backed by an assurance process, and price that liability into its fees. The practical form this takes is a warranty: the platform warrants that a data field meets a stated accuracy standard under stated conditions, and if it does not, the platform indemnifies the user up to a capped amount. The cap is the negotiation. Set it too low and the warranty is worthless; set it too high and the platform cannot obtain insurance. The insurers who write errors-and-omissions cover for data vendors are the invisible party in this negotiation, and their underwriting standards, not the platform's engineering, will determine how much liability the platform can absorb.

The third is privacy and commercial confidentiality. Firm-level environmental data are commercially sensitive: emissions intensity reveals production processes, and water use reveals capacity utilization. Farm-level data are personally identifiable. The platform must therefore support aggregation and differential privacy where the data are sensitive, and it must be explicit about the re-identification risk. The Chinese experience with big data analytics is relevant here: the gains from data aggregation are real, but as the digital finance literature has emphasized, data practices built on inadequate consent are not sustainable. A platform that cannot assure firms that their process data will not reach competitors will not receive accurate data, and inaccurate data are worse than no data. The design rule that follows is that the platform should expose derived indicators—emissions intensity relative to sector, trajectory over time—rather than raw process parameters, and it should allow a firm to see exactly what a competitor or a regulator can see about it.

The fourth is standards and interoperability. A platform that uses proprietary schemas and closed APIs creates a new silo. The standards that matter are the taxonomy (which defines the categories), the accounting standard (which defines the boundaries), and the data schema (which defines the fields). Alignment with the EU Taxonomy, the ISSB sustainability disclosure standards, and the PCAF financed-emissions standard is not optional for a platform that wants to serve internationally active institutions. The governance question is who has the authority to update the standards as they evolve, and how conflicts between standards are resolved. The conflicts are not hypothetical: the EU Taxonomy's do-no-significant-harm principle and China's Green Bond Endorsed Projects Catalogue classify the same coal-efficiency retrofit differently, and a platform serving both markets must either carry both classifications or choose one and explain the choice to the other's regulator.

12.3.3 The Chinese Case: Digital Finance Infrastructure as a Precedent

China offers the most instructive precedent for platform building, and also the most instructive cautionary tale.

The precedent is the speed at which digital payment and credit infrastructure was built. The PKU-DFIIC index documents the spread of digital financial services across provinces, cities, and counties over roughly a decade, and the associated research—by Huang and colleagues, and by the BIS group of Frost, Gambacorta, Huang, Shin, and Zbinden on BigTech and the changing structure of financial intermediation—explains the mechanism. The platform economics of payments created a data asset that could be used for credit scoring, and the credit scoring extended financial access to populations that bureau-based systems had excluded. The speed is the point: the index rose from 100 to 300 on its scale in under eight years, a diffusion rate that no traditional banking infrastructure had achieved.

The cautionary tale is the regulatory reckoning that followed. The Chinese digital finance literature describes a transition from a spontaneous development path to a more ordered one—what some of the Chinese writing calls the 2.0 era—in which the division of labor between finance and technology is clarified, technology firms provide technical solutions to financial institutions, and financial institutions remain responsible for financial services. The regulatory coverage became comprehensive, and the authorities paired it with regulatory sandboxes and controlled liberalization. The prohibition of cryptocurrency trading and initial coin offerings is the sharpest example, and the subsequent realization that Chinese institutions had no way to compete with a global stablecoin like Libra is the sharpest illustration of the cost of prohibition without a substitute.

The lesson for green finance data platforms is that the regulatory framework and the technical platform must be designed together. A platform built without a regulatory mandate will not receive mandatory data. A regulatory mandate imposed without a platform will produce reporting burdens without usable data. The Chinese green finance system has both a mandate—the Green Credit Statistics System—and a growing platform infrastructure, and the integration between them is the live policy question. The specific gap is visible in the statistics: the Green Credit Statistics System collects 21 indicators from reporting banks, and the indicators are self-reported at the entity level. The platform infrastructure that could verify them at the facility level exists in fragments—the CEMS network, the satellite archives, the provincial environmental registries—but no single institution has the mandate to fuse them. The Tangshan dispute was a symptom of that gap, and the fix is an institutional mandate, not a better model.

12.3.4 Cost, Benefit, and Who Pays

The final governance question is who pays for the platform, and this is where the economics get uncomfortable.

The benefits of a green finance data platform are diffuse. Every bank benefits from better data, but no single bank has an incentive to pay for the substrate that produces them, because the substrate is a public good. The standard remedies for public goods are public funding, mandatory contribution, or club goods with exclusion.

Public funding is the cleanest in principle and the hardest in fiscal practice. A green data platform competes with other public investments, and its benefits are hard to demonstrate ex ante. Mandatory contribution—requiring regulated financial institutions to fund the platform through a levy—aligns the cost with the beneficiaries and is the model used for deposit insurance and, in some jurisdictions, for credit bureaus. Club goods with exclusion—a consortium that restricts access to members—solve the free-rider problem but limit reach, which defeats the purpose for small borrowers.

My own view, based on the precedents, is that the substrate should be publicly funded and openly accessible, while the value-added services—models, dashboards, assurance—should be competitively supplied. The substrate is the identifier layer, the ingestion layer, and the basic quality layer. The services are the model layer and the access layer. This division has a clear economic logic: the substrate has natural monopoly characteristics and the services do not. It also has a clear political logic: public funding of the substrate is defensible as infrastructure, in the same category as a statistical agency or a geodetic survey, while public funding of the services is not.

The cost of the substrate is not large relative to the financial flows it supports. A national green finance data platform with entity resolution, satellite ingestion, and quality assurance might cost in the tens of millions of dollars per year to operate, against green finance flows in the hundreds of billions. The constraint is not the money. It is the institutional capacity to run it and the political will to make data submission mandatory. The comparison that makes the case is with the geodetic survey: no one argues that a national mapping agency should be funded by the firms that use maps, and no one argues that map data should be proprietary. The green data substrate belongs in the same category, and the reason it is not there yet is that the category is new.

12.3.5 What the Platform Does Not Solve

I want to close this section by being explicit about the limits, because the case for digital infrastructure is strong enough that it invites overclaiming.

The platform does not solve the additionality problem. It can measure whether a project reduced emissions relative to a baseline; it cannot determine whether the reduction would have happened without the finance. Additionality is a counterfactual, and counterfactuals require assumptions that no amount of data resolves. The best the platform can do is make the assumptions explicit and test their sensitivity. The forest carbon audits cited earlier are the cautionary case: the ledger recorded the credit faithfully, and the credit was not additional, because the baseline was wrong. Better measurement of the baseline would have caught the error, but no measurement resolves the counterfactual.

The platform does not solve the transition problem for brown firms. A steel mill that is improving but still far from green needs finance to continue improving, and a taxonomy built around thresholds may exclude it. The data platform can measure the trajectory, but the taxonomy and the pricing must decide what to do with it. That is a policy choice, not a data problem. The Tangshan mill is the case in point: its trajectory was improving, its level was above the threshold, and the platform could show both facts. What to do with them was the committee's decision.

The platform does not solve the incentive problem at the point of origination. A loan officer who is compensated on volume will use whatever data are convenient, and a platform that adds friction will be circumvented. The incentive design of Chapter 11 and the contract design of the preceding chapter are complementary to the platform, not substitutes for it.

And the platform does not solve the problem of trust in the underlying institutions. A perfectly engineered data pipeline that draws on CEMS readings from a system with documented tampering will produce precise, timely, and wrong numbers. The quality layer can flag discrepancies, but it cannot manufacture the audit capacity that makes the underlying measurement credible. That capacity is institutional, and it is built by enforcement, not by software. The 14 percent tampering rate in the 2019 audit is the number to keep in view: a platform that ingests tampered data at hourly intervals is a faster conduit for wrong numbers, not a solution to them.

12.4 Conclusion: The Data Substrate of the Three Pillars

The argument of this chapter can be stated in one sentence: the three-pillar framework of this book—rating-based identification, insurance-based risk buffering, sci-tech finance catalysis—rests on a data substrate that does not yet exist at the scale required, and building it is a governance problem more than a technology problem.

The evidence for the gap is in the aggregates. A green loan book at 5 percent of total lending, a green bond market at a few hundred billion dollars of annual issuance, and an environmental data infrastructure that remains largely self-reported: these are the marks of a financial system that has not yet built the measurement capacity its mandates require. The evidence for the technological feasibility is in the sectors where it has been built—satellite-based crop monitoring, alternative-data credit scoring, metered emissions in regulated industries. The gap between the feasible and the actual is institutional.

Three implications follow for the framework.

First, the rating pillar of Chapter 3 and the risk models of Chapter 4 should be designed as consumers of platform data, not as standalone systems that each institution builds for itself. The environmental variables that enter a default model should come from a shared, quality-scored source, so that differences in assessed risk reflect differences in the borrower rather than differences in the data vendor.

Second, the insurance pillar of Chapters 6 through 8 should be designed around digital index measurement from the outset. The administrative cost per policy is the binding constraint on reaching smallholders, and remote sensing and automated claims processing are what bring that cost down. A weather index insurance scheme that still requires a field adjuster per claim has not escaped the cost structure of traditional indemnity insurance; it has merely relabeled it.

Third, the catalysis pillar of Chapters 9 and 10 requires a different data asset from the first two. Rating and insurance need data on the environmental performance of established firms and farms. Catalysis needs data on the technology pipeline: which firms hold which patents, which projects are in which stage of development, which venture funds are investing in which technologies at which valuations. The platform architecture should accommodate this third data domain, but it should not pretend that the same fields serve all three. A patent database and an emissions registry are different instruments, and the governance that suits one may not suit the other.

Return to the steel mill in Tangshan. In a system with an integrated platform, the loan officer's screen shows the mill's CEMS readings for the past eight quarters, cross-checked against its reported emissions and against satellite thermal anomalies at the plant's coordinates. The quality score on the emissions figure is high, because the CEMS is audited and the three sources agree within tolerance. The mill's emissions intensity is in the second quartile of its sector, improving at about 3 percent a year. The platform's scenario module shows the loan's probability of default under a carbon price path consistent with the jurisdiction's stated target, and the increase relative to the base case is modest but not zero. The green designation rests on the mill's measured trajectory, and the loan is priced off that reading.

None of this requires a technology that does not exist. It requires an identifier that resolves the mill across four registries, an ingestion pipeline that pulls the CEMS data and the satellite data on the same schedule, a quality layer that does the cross-check, a model layer that maps the emissions trajectory into a default probability, and an access layer that shows the loan officer the result without exposing the mill's process data to its competitors. Each of these components has been built somewhere. The work is to build them together, under a governance arrangement that makes the data trustworthy and the platform sustainable.

The financial system has spent a decade learning to talk about green finance. The next decade is about learning to measure it. The measurement is not an adjunct to the financing; it is the condition under which the financing can be called green at all. A loan that is labeled green on the basis of a questionnaire is a claim. A loan that is priced green on the basis of a verified emissions trajectory is a fact about the world. The difference between the two is the data substrate.

The next step is concrete and institutional. A green finance data platform needs a mandating authority, a funding line, and a first use case before it needs a technology stack, and the three can be assembled in a single budget cycle if the political will exists. The authority is the central bank or the financial regulator, acting under the green finance mandate it already holds. The funding line is a public appropriation for the substrate, sized against the geodetic-survey precedent rather than against the private-sector analytics market. The first use case is the one this chapter opened with: entity resolution and emissions verification for the industrial sectors that the Green Credit Statistics System already covers, delivered to the reporting banks as a shared utility rather than as a per-institution build. The Tangshan loan committee should be able to see the mill's verified trajectory on the same screen as its financial statements. Until it can, the green label will remain a claim, and the three pillars will rest on a substrate that is not there.