International Multilateral Public Climate Finance Dataset from 2000 to 2023

Abstract

International multilateral public climate finance plays a key role in the global climate financing landscape, yet the availability of a comprehensive, publicly accessible, and transparent database remains inadequate. This study establishes an international multilateral public climate finance database covering the period from 2000 to 2023. The primary objective is to identify and track climate-related projects and their funding flows supported by major multilateral development finance institutions (MDFIs) and multilateral climate funds (MCFs). Using a machine learning model MLCF-BERT (Multilateral Climate Finance BERT), we extracted and classified climate mitigation and adaptation-related projects from project-level data, creating a climate finance database containing detailed information on funding sources, institutions, project types, recipient countries, financing amounts, and financing types. This database provides valuable insights into the distribution and trends of international climate finance, offering data support for climate finance policies and decision-making.

Background & Summary

Providing financial resources to reduce net greenhouse gas emissions and enhance climate adaptation capacity is a critical enabler of the low-carbon transition^1,2. Despite increasing commitments from both public and private sectors, the scale of investment required to address climate change far exceeds the currently available financial resources^3,4,5. According to estimates by the Climate Policy Initiative (CPI), achieving the 1.5 °C scenario will require an average of $7.4 trillion annually in climate finance by 2030, representing a fivefold increase from current levels⁶. Moreover, the majority of this funding will need to be directed to developing countries, which are most vulnerable to the impacts of climate change^7,8.

International public multilateral climate finance is a key component of the global climate finance architecture, providing essential financial support for mitigation and adaptation actions, particularly to emerging markets and developing economies (EMDEs) as well as least-developed countries⁹. Currently, many developing countries face constrained public budgets due to inflationary pressures and the post-pandemic economic recovery, leaving limited public funds available for low-carbon transitions and necessitating international support^10,11. The significance of multilateral funds lies in their ability to pool resources from multiple donor countries, thereby reducing transaction costs and enhancing the efficiency of resource allocation¹². Furthermore, these funds often act as catalysts, leveraging additional private and public investment in climate projects—an aspect that is especially critical for financing large-scale renewable energy and infrastructure projects in developing countries^13,14. Notably, international public actors such as International Development Finance Institutions have been highlighted as key players in overcoming barriers to renewable energy investment in developing countries¹⁵. By providing climate finance and mobilizing private capital for renewable energy projects at low financing costs, these institutions are expected to help unlock the renewable energy potential in developing nations¹⁶.

The sources of international public multilateral climate finance primarily include Multilateral Development Financial Institutions (MDFIs) and Multilateral Climate Funds (MCFs)^17,18. MDFIs are international financial organizations jointly established by two or more sovereign countries, with the primary objective of promoting the economic, social, and sustainable development of developing countries through the provision of concessional financing, technical assistance, and policy guidance. Such joint mechanisms not only facilitate the pooling of resources and the sharing of development risks among countries but also offer an effective governance model for delivering transnational public goods and advancing international cooperation. Through loans, grants, and technical assistance, MDFIs support a variety of development projects across fields such as infrastructure, education, healthcare, and environmental protection¹⁹. MDFIs have been at the core of the global development and climate finance agenda²⁰. In 2022, development financial institutions accounted for 86% of tracked adaptation financing, with MDFIs contributing 34% of the total²¹. The G20 has called for reforms to the capital adequacy frameworks of multilateral development banks, which could potentially increase lending capacity by $357 billion over the next decade²². Given the climate-related missions and contributions of multilateral development banks, their importance is likely to grow further if calls for MDB capitalization and expansion are realized, representing a significant opportunity to scale up global climate finance²³. MCFs are multilateral climate funds established with contributions from multiple countries and managed by multilateral development institutions or specialized international agencies. These funds play a critical role in supporting developing countries in transitioning toward low-emission and climate-resilient futures. MCFs typically provide critical financing for the deployment of low-carbon technologies and the development of climate-resilient infrastructure in developing countries, primarily in the form of grants or concessional loans. Beyond financial support, MCFs also facilitate low-carbon technology transfer and capacity building, helping developing countries bridge the technology gap, enhance climate resilience, and promote inclusive development⁹.

Several studies and organizations have compiled and analyzed international public multilateral climate finance (Table 1). For example, Climate Funds Update (CFU) is an independent website that provides information and data on multilateral climate finance initiatives. It summarizes the financial flows, project distribution, and implementation status of major MCFs⁹. In recent years, some multilateral development banks (MDBs) have jointly issued the Joint Report on Multilateral Development Banks’ Climate Finance, disclosing their contributions to international climate finance²⁴. Some organizations also aggregate data from MDFIs and MCFs. For example, the UNFCCC Biennial Assessment and Overview of Climate Finance Flows is a biennial report prepared by the Standing Committee on Finance (SCF) under the UNFCCC. Its purpose is to assess global climate finance flows, enhance transparency, and monitor progress toward financial targets²⁵. The Organization for Economic Cooperation and Development (OECD) provides important statistics on development aid flows to developing countries through its Official Development Assistance (ODA) data. The Rio Markers, part of this dataset, serve as a classification mechanism to identify and track climate-related financial flows within the development aid²⁶. The CPI also regularly publishes the Global Landscape of Climate Finance, which provides a comprehensive assessment of global climate finance flows⁶.

Table 1 Summary of Existing Statistics on International Multilateral Public Climate Finance.

While progress has been made in the statistical efforts concerning international public multilateral climate finance, significant challenges remain, affecting the accuracy and comprehensiveness of the data. These limitations hinder effective monitoring and analysis of global climate finance flows. First, there are discrepancies in the definitions and statistical scope of climate finance among institutions and researchers. For instance, some statistics only include funding directly related to climate change mitigation and adaptation, while others may also incorporate funding for sustainable development, leading to inconsistencies in data standards. For example, existing studies have questioned the accuracy of ODA data, pointing out that climate finance data reported under the Rio Markers may be overestimated²⁷. Additionally, different statistical agencies employ varying metrics and methodologies for data collection and reporting, making it challenging to aggregate and integrate the datasets²⁸. Second, a lack of timeliness is a critical shortcoming of these databases. Most climate finance reports are published annually or biennially, resulting in a time lag that fails to reflect the latest trends and shifts in climate finance flows. Another issue is the insufficient differentiation of funding types in existing statistics. Most focus solely on the sources and aggregate amounts of funds, overlooking distinctions such as grants, concessional loans, guarantees, and equity investments. This hinders an accurate analysis of the current structure of climate finance. Finally, current statistics often emphasize the contributions of fund providers without adequately tracking how funds are allocated to projects, regions, and specific beneficiaries. This lack of detailed information undermines transparency and makes it difficult to evaluate the effectiveness of fund allocation.

To address these important research gaps, this study innovatively constructs the MLCF-BERT (Multilateral Climate Finance BERT) machine learning model and develops an international multilateral public climate finance database (2000–2023) based on project-level data. Leveraging project-level data from major global MDFIs and MCFs, the model systematically identifies, classifies, and tracks climate change-related projects. By applying a standardized methodology across different institutions, MLCF-BERT effectively reconciles discrepancies in the statistical scope of climate finance accounting and facilitates the aggregation and integration of data from diverse sources. The model also extracts detailed project-level information, including funding sources, implementing institutions, project types, financing amounts, regions, countries, project years, and financing instruments. Crucially, the model supports real-time updates of climate finance data based on newly disclosed project information, thereby overcoming the temporal limitations associated with reliance on annual institutional reports. This comprehensive database enables a deeper understanding of the distribution and flow patterns of international public multilateral climate finance, offering robust data support for climate finance policy and decision-making.

The database constructed in this study makes several key contributions to the existing literature. First, our database covers a broad range of institutions and time periods, encompassing data from major global MDFIs and MCFs over 24 years (2000–2023). The coverage of MDFIs is broader than that in the reports of multilateral development banks, and the time span is longer than that in most other data sources. Second, our methodology allows for consistent climate finance estimation across different institutions and funds, addressing the issue of comparability and aggregation across different reports. Third, we differentiate between various types of financing, including grants, loans, guarantees, and equity. Additionally, we focus on the specific countries and sectors receiving the funds, beyond just the fund providers. Finally, our project identification approach enables the climate finance data to be updated in real-time based on publicly available project information, ensuring both the timeliness and accuracy of capturing the dynamics of climate finance development.

Methods

This study aims to construct a global public multilateral climate finance database covering the period from 2000 to 2023. We selected major MDFIs and MCFs worldwide. The selection of MDFIs was based on the Public Development Finance Institutions Database (INSE-AFD Joint Database) built by the Peking University New Structural Economics Institute and the French Development Agency (AFD). We considered multilateral development finance institutions classified as super-large, large, and medium-sized in their database. We excluded institutions for which project information could not be retrieved from their websites and included relatively larger small-sized institutions. In total, 14 MDFIs were included in our study, representing 90% of the total assets of all MDFIs in their database. For the MCFs, we referred to the global climate finance framework established by Climate Funds Update, which includes major UNFCCC and non-UNFCCC funding mechanisms. The UNFCCC funding mechanisms include the Global Environment Facility (GEF) and its associated funds, the Green Climate Fund (GCF), and the Adaptation Fund (AF). Non-UNFCCC funding mechanisms primarily include the Climate Investment Funds (CIF) and its various funds. The MDFIs and MCFs considered in this study are listed in Table 2, with their source URLs and the number of included projects provided in Supplementary Table 1.

Table 2 MDFIs and MCFs Included in the Study.

The process of constructing the database involves four main steps (Fig. 1): (1) Downloading or extracting project information from the official websites of various MDFIs and MCFs, and cleaning the project description texts; (2) Classifying the climate projects; (3) Training a machine learning model to identify and classify projects based on their descriptive text; (4) Using the trained model to identify and classify climate projects within the MDFIs and MCFs, calculating the financing amount for each climate project, and merging this data with other project information to generate the database.

Fig. 1

Research framework for constructing the global public multilateral climate finance database.

Step1: Data collection and cleaning

We first visited the official websites of various institutions and funds to search for project information lists. Some websites allow for direct downloading of bundled information, while many others do not provide direct download options. Additionally, most of the downloadable data does not fully meet our requirements, necessitating manual supplementation. To address this, we used the Selenium tool in Python. Selenium is a popular Python library for automating web interactions, which allows us to locate webpage elements and extract relevant project information. The collected project information includes at least the project name, commitment year, project description, and financing details; any additional available information was also collected.

Next, we filtered projects with commitment years from 2000 to 2023. The collected data then underwent data cleaning, primarily focusing on project names and descriptions. First, we translated non-English project texts into English using the Google Translate API in Python. This tool provides high-quality machine translation and supports fast conversion between multiple languages. The text was then standardized to lowercase, and unnecessary characters were removed. This step laid the foundation for subsequent model training.

Through the above data collection and cleaning process, we obtained a preliminary project information aggregation database, which includes 68,892 project entries from MDFIs and 7,158 project entries from MCFs, totaling 76,050 data points.

Step2: Climate Finance Project Categorization

Climate projects are primarily divided into two categories: mitigation and adaptation. Climate change mitigation refers to actions taken to reduce greenhouse gas emissions or increase carbon sinks in order to lower the concentration of greenhouse gases in the atmosphere, thereby slowing the rate of global warming. Climate change adaptation refers to measures taken to respond to the impacts of climate change that have already occurred or are expected to occur, aiming to reduce negative effects on natural ecosystems and socio-economic systems while also leveraging potential opportunities brought by climate change.

Climate projects were classified based on existing literature on the topic, with the classification carried out at the project level. A detailed categorization is provided in Table 3. Within the mitigation category, we further subdivided projects into 9 types, including various renewable energy projects, energy efficiency improvement projects, and carbon dioxide removal technology projects. Projects that could not be clearly classified into these three types were categorized as “Other Mitigation Projects.” Adaptation projects include several categories such as disaster risk reduction, coastal management, and ecosystem-based adaptation, among others. However, due to the higher difficulty in model identification, we did not further subdivide these categories.

Table 3 MDFIs and MCFs Included in the Study.

Additionally, we identified environmental projects that may have positive climate benefits, including biodiversity, sustainable land use, and water management projects. However, it is important to note that these projects were not included in the calculation of climate finance volumes in the subsequent analysis. It is worth noting that we classified all projects into only one category, even though some projects may have multiple positive impacts, such as co-benefits of mitigation and adaptation. This approach facilitates the statistical analysis of climate finance amounts, and the results from model training further indicate that this method does not introduce significant bias. To provide greater clarity on how the classification framework is applied in practice, Table 4 presents five representative examples from our annotated dataset.

Table 4 An example of the classification of climate projects in the annotation set.

Step3: Model training and validation

Next, we built a machine learning model, MLCF-BERT, for the identification and classification of climate projects. To provide training samples for the MLCF-BERT model, we first performed stratified sampling based on the number of projects in different institutions, resulting in 1,600 projects, which were manually classified and annotated. These projects were selected from categories that are likely related to climate change within these institutions, such as agriculture, energy, electricity, transportation, and environmental protection. This was done to increase the concentration of annotated samples that could be used for training a multi-label classifier. We then manually added 50 projects from underrepresented categories to improve the classification accuracy for these categories. These 1,650 projects formed our manually annotated dataset.

Our MLCF-BERT model is built upon the ClimateBERT model (https://huggingface.co/climatebert/distilroberta-base-climate-f). ClimateBERT is a pre-trained language model specifically designed for climate change-related tasks. It is based on a variant of the BERT (Bidirectional Encoder Representations from Transformers) architecture called DistilRoBERTa. The model was fine-tuned and trained using a specialized corpus of over 2 million climate-related texts from the fields of climate change and environmental science, aimed at enhancing the understanding and response capabilities regarding climate change issues.

Following the approach of Toetzke et al.²⁷, we also divided the model into a relevance classifier and a multi-label classifier. The relevance classifier assesses the relationship of projects to mitigation, adaptation, and the environment, while the multi-label classifier assigns projects that are deemed relevant by the first classifier to the most appropriate climate finance category. Both classifiers were trained and validated separately. We randomly split the training data into a training set, a validation set, and a test set. For the relevance classifier, the training set contained 85% of the data, with the validation and test sets each comprising 7.5%. For the multi-label classifier, the training set accounted for 65%, with the validation and test set each representing 17.5%. The differential data allocation for the two classifiers was to ensure that the test sets of the multi-label classifier had a similar size to those of the relevance classifier. For both models, we used the AdamW optimizer, iterated 50 and 100 times respectively, and employed cross-entropy loss for model validation.

Step4: Project classification and database construction

After the model training was completed, we used the trained MLCF-BERT model to identify and classify other projects. It should be noted that the projects from multilateral development finance institutions underwent both the relevance classifier and the multi-label classifier. However, the treatment of projects from international MCFs is different. The projects of the GEF are tagged with climate-related labels, so we consider GEF projects in “Focal Areas” related to climate change as climate projects. For the projects of GCF, AF, and CIF, we uniformly classify them as climate-related and use only the multi-label classifier to assign them to different climate project categories. Subsequently, we aggregated the funding amounts of the identified climate projects and merged them with other project information to compile the global public multilateral climate finance database.

Data Records

The dataset is available at figshare²⁹. We provide three datasets as part of the global multilateral public climate finance database, covering the period from 2000 to 2023. These datasets are available in Excel format to ensure compatibility with common analytical tools.

• Overall Climate Finance Dataset (Climate_finance_data_2000_2023.xlsx): This dataset contains information on global public multilateral climate finance from 2000 to 2023. It includes details such as the year, climate project type (e.g., mitigation, adaptation, and cross-cutting), total climate finance amounts, and disaggregated amounts by funding type (e.g., loans, grants, etc.).

• Recipient Country Dataset (Climate_finance_data_by_recipient_country folder): This dataset provides information on global public multilateral climate finance received by individual countries from 2000 to 2023. Each file in this folder corresponds to a specific country and is stored in Excel format. The data include the country name, year, climate project type, total climate finance amounts, and disaggregated amounts by funding type.

• Institution-Level Dataset (Climate_finance_data_by_institution folder): This dataset focuses on the contributions of individual MDFIs and MCFs from 2000 to 2023. Each file in this folder corresponds to a specific institution and is stored in Excel format. It includes the institution name, year, climate project type, total climate finance amounts, and disaggregated amounts by funding type.

Technical Validation

Model validation

We report the performance results of both the relevance classifier and the multi-label classifier on the test set. Table 5 shows the performance of the relevance classifier, with an overall accuracy of 94% on the test data. Table 6 presents the performance results of the multi-label classifier, with an overall accuracy of 90%. For the multi-label classifier, we further consolidated the categories into three main groups: mitigation, adaptation, and environment, and found that the accuracy reached 93%. This indicates that most classification errors occurred within these three main categories.

Table 5 Performance results of the relevance classifier.

Table 6 Performance results of the multi-label classifier.

We report the results of the model’s precision, recall, and f1-score. Precision represents the proportion of samples predicted by the model to belong to a certain category that actually belong to that category. Recall represents the proportion of samples that actually belong to a certain category and are correctly predicted as such by the model. The f1-score is the harmonic mean of precision and recall, used to comprehensively evaluate the model’s accuracy and recall performance.

Manual checking

To further validate the accuracy of the MLCF-BERT model’s predictions, we performed a manual verification of the classification results by randomly selecting 500 projects based on different institution types through stratified random sampling. These projects include 106 adaptation projects, 312 mitigation projects, and 82 environmental projects. Table 7 illustrates the accuracy of predictions for different categories. Among these, the accuracy of predicting adaptation projects is 0.87, while the accuracy for environmental projects is 0.96. The overall accuracy for mitigating projects is 0.88, with category-specific accuracies ranging from 0.75 to 1.00, with the lowest accuracy observed for carbon removal technology projects. Overall, the model performs well, with most prediction errors occurring within major categories.

Table 7 Comparison of manual checking and machine learning prediction results.

Comparison with existing datasets

We compare the dataset constructed in this study with existing data to assess its accuracy and consistency. Specifically, we selected data from the annual Global Landscape of Climate Finance report published by CPI (https://www.climatepolicyinitiative.org/publication/global-landscape-of-climate-finance-2023/) and the Joint Report on Multilateral Development Bank Climate Finance for comparison (https://www.eib.org/en/publications/20240150-2023-joint-report-on-multilateral-development-banks-climate-finance)^21,24. One reason for selecting these two datasets is that they provide data with similar coverage to ours. The CPI report includes climate finance data from MDFIs. Other datasets either present total climate finance or only differentiate between public and private funding, which do not align with the scope of our dataset and are therefore not suitable for comparison. Another reason for our choice is that, although our dataset includes both MDFI and MCF data, we assume that all MCF projects are climate-related (with GEF projects having climate tags), making the comparison with MDFI data the primary focus. Consequently, we selected the Joint Report on Multilateral Development Bank Climate Finance for this purpose.

First, we compare our dataset with the data from The Global Landscape of Climate Finance published annually by CPI (Fig. 2). We compare the climate finance data for “Multilateral DFIs” from the CPI report with the MDFI data in our dataset and find that the trends by year are largely consistent, with a calculated R² of 0.51. The data in this study’s dataset is lower than that in the CPI report. One possible reason for this discrepancy is that CPI collects data from some smaller development finance institutions through surveys, which are not included in our database. Additionally, the CPI report does not publicly disclose the data for the project portfolio, so the accuracy of their climate finance estimates cannot be fully guaranteed, which may also account for the differences in the data.

Fig. 2

Comparison of the Dataset with the CPI Report.

We also compare the climate finance data by institution in our dataset with the data from the Joint Report on Multilateral Development Bank Climate Finance (Fig. 3). It can be observed that the trends for most institutions in the two datasets are consistent, especially after 2016. The data for ADB, AIIB, and EBRD shows a high degree of overlap, while the climate finance data trends for AfDB, EIB, IDB, and WB are consistent, although the climate finance data in the multilateral development bank report is higher than the estimates in this study. Figure 3 highlights two notable crossover points. The first is the significant increase in climate finance reported by EIB in the 2019 Joint MDB Report, which exceeds the figures in our dataset. This shift can be attributed to a methodological change in the 2019 Joint MDB Reports, which expanded the reporting scope to include all economies where MDBs operate, rather than only developing and emerging economies. As a result, both EIB and EBRD reported significantly higher climate finance. The second is the notable increase in IDB’s reported data in 2017. This is due to a change beginning in 2016 when the IDB Group started including private-sector climate finance activities undertaken by the Inter-American Investment Corporation (IIC), which had previously been excluded. The inclusion of these activities led to higher reported amounts in the Joint MDB Reports compared to those in our dataset. Some studies on these institutions indicate that the data in the Multilateral Development Bank joint report is self-reported and has not been independently verified by a third party. Furthermore, multilateral development banks do not disclose the details of how climate finance is assessed for each individual project, and the lack of transparency and information disclosure diminishes credibility³⁰. In addition, the format and data of the annual joint report often change, making it challenging to assess trends over a period of time. This study also makes an important contribution to enhancing the credibility of multilateral climate finance data to some extent.

Fig. 3

Comparison of the Dataset with the Multilateral Development Bank Joint Report Data by Institution.

Usage Notes

Applications and relevance of the dataset

The datasets constructed in this study provide a robust foundation for analyzing trends in global multilateral public climate finance, evaluating climate finance policies, studying the behavior of MDFIs and MCFs, and conducting future scenario analyses. This study offers comprehensive data on global multilateral public climate finance from 2000 to 2023, including detailed information categorized by recipient country, institution, and project type. These datasets enable the analysis of temporal trends, geographic distributions, and financial flows, providing essential support for understanding the evolution of global climate finance. By examining the allocation and utilization of funds, these datasets can assess the effectiveness of climate finance policies. For instance, they allow researchers to evaluate whether the distribution of funding for mitigation and adaptation projects aligns with the climate goals of recipient countries and to analyze the actual impacts of different funding types, such as loans and grants. The datasets also include the financial contributions of MDFIs and MCFs, facilitating the study of their respective roles and priorities in global climate finance. For example, they can help identify whether certain institutions focus more on mitigation efforts or prioritize adaptation investments. Furthermore, the datasets support the construction of future climate finance scenarios, such as predicting the potential impacts of various financing models on achieving the goals of the Paris Agreement. Researchers can use these datasets to propose policy recommendations aimed at optimizing funding allocation strategies and improving financing mechanisms. These datasets are a valuable resource for policymakers, researchers, and practitioners seeking to advance the understanding and implementation of effective climate finance strategies worldwide.

Uncertainties and limitations

While this dataset provides comprehensive insights into global multilateral public climate finance, users should be aware of several uncertainties and limitations. First, the dataset primarily focuses on large MDFIs and major MCFs, with data from smaller institutions yet to be included. Future research, as data availability permits, could consider expanding the scope to include more financial institutions, thus offering a more complete picture of global climate finance. Second, some MCFs lack detailed project descriptions, and therefore, all projects within these funds are considered climate-related in the current database (with GEF projects labeled as climate-tagged). Future work could incorporate project descriptions to provide a more accurate classification of projects. Third, due to the difficulty in isolating climate-related activities within broader projects, all funding amounts associated with climate-related projects are classified under climate finance. This approach is consistent with most current reporting practices. Additionally, since data from multilateral institutions do not consistently provide detailed information on sub-projects and their associated funding, projects in the database are classified at the project level. This classification method reflects the finest granularity of the available data, though it does not rule out the possibility that some projects may include sub-projects targeting different climate goals. If data becomes available, this aspect could be considered in future research. Fourth, climate-related projects are identified using a machine learning model, which introduces classification uncertainty, especially for projects with multiple or unclear objectives. To avoid potential misinterpretation, only aggregated results by institution and country are shared, rather than project-level classifications. Lastly, the dataset includes amounts and years of commitments made by the institutions, which do not directly represent the actual disbursement of funds, meaning the data reflects commitments rather than the realization of funding.