Improving economic impact assessment of climate change with machine learningDownload PDF Download PDF CommentOpen accessPublished: 25 June 2026Anton Orlov ORCID: orcid.org/0000-0003-3670-597X1 &Jana Sillmann ORCID: orcid.org/0000-0002-0219-53451,2 Nature Communications volume 17, Article number: 5574 (2026) Cite this articleSubjectsEconomicsEnvironmental economicsHybrid modelling using machine learning can help improve the robustness and comprehensiveness of economic impact assessments of climate change by combining data with theory, bridging silos across social, economic, and financial systems, and linking micro- and macroeconomic scales.Current economic modelling approachesThe IPCC’s Sixth Assessment Report shows that estimates of the economic impacts of climate change vary widely across studies due to methodological differences and underlying uncertainties in economic impact modelling1,2. While estimates of economic losses from climate change tend to increase over time3, persistent methodological differences continue to drive substantial variation across studies4. The forthcoming Seventh Assessment Report will include updated scenarios generated using the emulator approach (i.e., statistical models approximating process-based models to enable rapid simulations) to provide more robust and comprehensive impact assessments, including economic impacts5. Although broader scenario coverage improves risk assessment, persistent uncertainties and limitations in economic modelling remain. Robust and comprehensive assessments of the economic impacts of climate change are essential to support informed mitigation and adaptation decisions.The economic impacts of climate change are typically assessed using two approaches: statistical models and structural models1. Statistical (econometric) models use historical data to estimate reduced-form relationships between climate variables (e.g., temperature and precipitation) and economic productivity (typically measured by gross domestic product, GDP), without explicitly modelling the underlying transmission mechanisms6,7. Structural models, such as computable general equilibrium (CGE) models, agent-based models (ABMs), and dynamic stochastic general equilibrium (DSGE) models, are analogous to process-based (or mechanistic) models. Among these, CGE models are most widely used in applied economic assessments of climate change. They represent economic behaviour based on theoretical principles (e.g., utility and profit maximisation and equilibrium) and depict economic interactions through systems of equations. In structural models, climate impacts are incorporated as shocks to production and demand through sector-specific exposure-response functions or derived from process-based impact models. While structural models explicitly represent the mechanisms through which climate change affects economic outcomes, they often rely on idealised assumptions about socioeconomic interactions and can be difficult to validate empirically. These two approaches produce substantially different estimates of macroeconomic climate costs (e.g., GDP impacts)4, underscoring the need to reconcile them to reduce uncertainty in estimates of economic impacts.Beyond reducing uncertainty in estimates of macroeconomic costs, more detailed and comprehensive impact assessments are needed that go beyond the aggregate GDP metric8. Higher spatial and temporal resolution in economic models helps to better capture direct impacts on households and firms. Broader cross-regional and cross-sectoral coverage is needed to capture indirect impacts propagating through global supply chains, capital flows, and migration9. Machine learning (ML) can help reconcile estimates of economic impacts from statistical and structural approaches while improving the granularity and comprehensiveness of impact assessments. Figure 1 illustrates how ML can improve the economic modelling of climate change impacts.Fig. 1Full size imagePathways through which machine learning can enhance the economic modelling of climate change impacts.The promise of hybrid economic modelling using machine learningThe trade-offs between statistical and structural models (broadly defined as data-driven vs. theory-driven approaches) are not unique to economics. They are also common across many fields, including those studying climate change impacts (e.g., crop modelling)10,11. Hybrid modelling approaches that combine data- and theory-driven models are increasingly being developed12, with ML playing a key role. Economic assessments of climate change can benefit from these advancements, as hybrid modelling is a promising and rapidly evolving approach.The application of ML in economics is not new13. It is widely used across subfields of economics (e.g., financial economics), leveraging rich micro-level data14. However, its use in economic impact assessments of climate change remains limited. A key challenge for comprehensive assessment across regions and sectors is the coarse spatial and temporal resolution of socioeconomic data (e.g., income and economic output). National accounts and sectoral statistics are typically aggregated due to reporting and accounting requirements and firm-level confidentiality. However, ongoing efforts to collect high-resolution economic data15 could open new opportunities for the application of ML methods. ML can enhance data generation by synthesising diverse data sources, imputing missing data, and processing socioeconomic data (e.g., household surveys, satellite observations, financial transactions)16,17,18. It can also be used to downscale aggregated economic impacts to local-level impacts19. While traditional statistical methods can handle such tasks, ML may outperform them in capturing nonlinear patterns in socioeconomic systems.Recently, ML has been used to approximate the outputs of structural models through surrogate models or emulators20,21, enabling faster simulations. Further research is needed to advance their development and applicability. However, emulators do not address key limitations of structural models (e.g., idealised assumptions and limited empirical validation). Structural models, particularly equilibrium-based models, often struggle to capture real-world complexities, such as institutional and market frictions, heterogeneous agents, information asymmetries, and interactions across economic, financial, and social systems. ABMs are designed to capture heterogeneity and disequilibrium dynamics but rely on stylised rule-based representations of agent behaviour and can be difficult to validate and are computationally demanding. As a result, emulators are constrained by the quality of the structural models they approximate.Hybrid economic modelling offers an alternative approach by combining ML and structural models in two complementary ways: (1) using ML to improve structural models and (2) using structural models to inform ML. First, ML can improve structural models by calibrating behavioural parameters (e.g., elasticities) or by replacing selected components with ML-based modules. These modules could be particularly useful for replacing components that are computationally intensive, data-rich but theory-poor, or less directly relevant to decision making. Second, structural insights can inform ML through Economics-Informed ML (EIML). This approach is analogous to Physics-Informed ML (PIML) or Neural Networks (PINN) models22, which embed physical principles into ML models as guiding structures. While PIML is increasingly used in climate-related applications23, its adoption in economic impact assessments of climate change remains largely unexplored. EIML would integrate key economic principles from structural models (e.g., profit maximising behaviour, and budget or resource constraints) as structural constraints (i.e., strict or weak constraints) into ML frameworks. This hybrid modelling approach could provide a balance between theory-driven and data-driven modelling tailored to decision-making needs, improving interpretability, generalisability, and scalability while maintaining the flexibility of purely data-driven approaches.Practical usefulness of hybrid economic models in decision makingStatistical models are grounded in data, less computationally intensive and often more user-friendly than structural models. However, statistical models are typically less interpretable and less comprehensive, as they estimate isolated reduced-form relationships rather than representing the underlying economic structure. Data-driven approaches remain limited in evaluating counterfactual policy interventions, for which no observational data exist. Structural models can simulate interventions outside observed data, but they are often based on simplifying assumptions. A key advantage of structural models is their ability to explicitly represent policy levers through underlying mechanisms. Hybrid models can combine empirical realism and causal structure to improve policy analysis. In hybrid modelling, the choice of which mechanisms should be explicitly represented through structural modelling, and which should be approximated using ML, depends on stakeholders’ needs. In principle, key behavioural mechanisms, economic constraints, and policy-relevant components should be modelled explicitly. Hybrid models could improve the robustness of policy simulations by combining causal structure with data-driven flexibility. Hybrid models may also provide more robust extrapolation under future climate and socioeconomic conditions than statistical models based solely on historical climate–economy relationships.Hybrid models offer several advantages for decision making. For example, in analysing supply chain impacts, EIML could incorporate structural features of supply chains (i.e., trade linkages and intermediate inputs), while allowing ML to capture behavioural responses of consumers and producers. Another application of EIML could be integrating DSGE models with ML. DSGE models, widely used by central banks, provide a strong theoretical foundation. Their structural elements (e.g., optimisation conditions and equilibrium constraints) could guide ML models. Alternatively, ML could approximate selected components of DSGE models to improve predictive accuracy and capture complex relationships, or be used to build emulators of these models24,25. Deep learning is increasingly used to approximate policy and value functions in DSGE models by solving Bellman and Euler equations26,27. ML could also enhance DSGE models by improving the representation of the financial sector, which can play a key role in amplifying physical and transition risks through cascading financial shocks28. The abundance of financial data can enable ML and hybrid models to address this gap.ML could help bridge macro- and microeconomic scales. Multi-scale assessments can help policymakers and financial institutions move beyond GDP to identify economic vulnerabilities and assess the cost-effectiveness of adaptation measures. For example, ML could improve the representation of climate-related income distributional effects (e.g., across income groups) in structural models (e.g., integrated assessment models (IAMs) and CGE models). While climate impacts have been shown to exacerbate inequality29, large-scale structural models often lack detailed representation of distributional outcomes. To address this limitation, structural models could provide an economy-wide perspective by capturing interactions across sectors and regions, while ML-based modules could complement them by translating aggregated economic outcomes into household- and firm-level impacts. ML can effectively capture rich, high-dimensional heterogeneity across households and firms, which is typically lost in aggregated structural models. This hybrid approach could incorporate feedback effects between macro- and microeconomic scales, enabling more precise local-level estimates.ML could also help bridge disciplinary silos. Climate, economic, and social systems are deeply interconnected through feedback loops with impacts on income, health, migration, and conflict30. These interactions are difficult to model mechanistically. ML could support more integrated analysis across these domains. For instance, ML-based modules for health, migration, conflict, and demographic change could be embedded within structural models to better capture cross‑sectoral and cross-regional socioeconomic dynamics.While hybrid models could enhance economic impact assessments, fundamental challenges remain. These include limited out-of-sample validity and deep structural uncertainties arising from complex interactions between climate, economy, institutions, demography, and technology. Compared to physical systems, socioeconomic systems are governed by less stable and more context-dependent relationships. Social system dynamics are difficult to model due to heterogeneity and behavioural variability. The structural insights underlying hybrid models are only as reliable as their ability to represent socioeconomic structures and dynamics. Thus, the capability of hybrid models to capture emergent behaviour is an important research avenue. Addressing these challenges requires interdisciplinary and transdisciplinary collaboration among climate scientists, economists, social scientists, data scientists, and practitioners, bringing together domain expertise, theoretical rigour, and data-driven methods.ReferencesO’Neil, B. & et al. Key risks across sectors and regions. In: Climate Change 2022: Impacts, Adaptation and Vulnerability. Contribution of Working Group II to the Sixth Assessment Report of the Intergovernmental Panel on Climate Change [H.-O. Pörtner et al. (Eds.)], pp. 2411–2538, https://doi.org/10.1017/9781009325844.025. (Cambridge University Press, 2022).Orlov, A. et al. Quantifying the cost of climate change impacts: limitations and frontiers of economic models. SSRN Scholarly Paper at https://doi.org/10.2139/ssrn.5493267 (2025).Tol, R. S. J. A meta-analysis of the total economic impact of climate change. Energy Policy 185, 113922 (2024).Article Google Scholar Morris, J. et al. Reconciling widely varying estimates of the global economic impacts from climate change. Nat. Clim. Change 15, 124–127 (2025).Article ADS Google Scholar van Vuuren, D. et al. The Scenario Model Intercomparison Project for CMIP7 (ScenarioMIP-CMIP7). EGUsphere 1, 38 (2025).Google Scholar Neal, T., Newell, B. R. & Pitman, A. Reconsidering the macroeconomic damage of severe warming. Environ. Res. Lett. 20, 044029 (2025).Article Google Scholar Newell, R. G., Prest, B. C. & Sexton, S. E. The GDP-Temperature relationship: Implications for climate change damages. J. Environ. Econ. Manag. 108, 102445 (2021).Article Google Scholar Jansen, A., Wang, R., Behrens, P. & Hoekstra, R. Beyond GDP: a review and conceptual framework for measuring sustainable and inclusive wellbeing. Lancet Planet. Health 8, e695–e705 (2024).Article PubMed Google Scholar van den Hurk, B. J. J. M. et al. Climate impact storylines for assessing socio-economic responses to remote events. Clim. Risk Manag. 40, 100500 (2023).Article Google Scholar Sweet, L. et al. Transdisciplinary coordination is essential for advancing agricultural modeling with machine learning. One Earth 8, 101233 (2025).Sillmann, J. et al. Combined impacts of climate and air pollution on human health and agricultural productivity. Environ. Res. Lett. 16, 093004 (2021).Article CAS Google Scholar Reichstein, M. et al. Deep learning and process understanding for data-driven Earth system science. Nature 566, 195–204 (2019).Article ADS CAS PubMed Google Scholar Athey, S. The impact of machine learning on economics. In The Economics of Artificial Intelligence: An Agenda (eds Agrawal, A., Gans, J. & Goldfarb, A.) https://doi.org/10.7208/chicago/9780226613475.003.0021 (University of Chicago Press, 2019).Zheng, Y., Xu, Z. & Xiao, A. Deep learning in economics: a systematic and critical review. Artif. Intell. Rev. 56, 9497–9539 (2023).Article Google Scholar Wenz, L., Carr, R. D., Kögel, N., Kotz, M. & Kalkuhl, M. DOSE—global data set of reported sub-national economic output. Sci. Data 10, 425 (2023).Article PubMed PubMed Central Google Scholar Rolf, E. et al. A generalizable and accessible approach to machine learning with global satellite imagery. Nat. Commun. 12, 4392 (2021).Article ADS CAS PubMed PubMed Central Google Scholar Lehnert, P., Niederberger, M., Backes-Gellner, U. & Bettinger, E. Proxying economic activity with daytime satellite imagery: filling data gaps across time and space. PNAS Nexus 2, pgad099 (2023).Article PubMed PubMed Central Google Scholar Moran, D., Giljum, S., Kanemoto, K. & Godar, J. From satellite to supply chain: new approaches connect earth observation to economic decisions. One Earth 3, 5–8 (2020).Article ADS Google Scholar Mikou, M., Vallet, A., Guivarch, C. & Makowski, D. High-resolution downscaling of disposable income in Europe using open-source data. Earths Future 13, e2024EF004576 (2025).Article ADS Google Scholar Takakura, J. et al. Reproducing complex simulations of economic impacts of climate change with lower-cost emulators. Geosci. Model Dev. 14, 3121–3140 (2021).Article ADS Google Scholar Dixon, P. B., Rimmer, M. T. & Schiffmann, F. Neural-network. approximation of reduced forms for cge models explained by elementary examples https://www.copsmodels.com/ftp/workpapr/g-348.pdf (2024).Karniadakis, G. E. et al. Physics-informed machine learning. Nat. Rev. Phys. 3, 422–440 (2021).Article Google Scholar Bracco, A. et al. Machine learning for the physics of climate. Nat. Rev. Phys. 7, 6–20 (2025).Article Google Scholar Nachane, D. M. & Chaubal, A. Machine Learning in Macroeconomics: Application to DSGE Models. in India’s Contemporary Macroeconomic Themes: Looking Beyond 2020 (eds Srivastava, D. K. & Shanmugam, K. R.) 521–541. https://doi.org/10.1007/978-981-99-5728-6_22 (Springer Nature, 2023).Sofianos, E., Betti, T., Papadimitriou, T., Barbier-Gauchard, A. & Gogas, P. Using DSGE and machine learning to forecast public debt for France. https://doi.org/10.1002/for.70144.Maliar, L., Maliar, S. & Winant, P. Deep learning for solving dynamic economic models. J. Monet. Econ. 122, 76–101 (2021).Article Google Scholar Fernández-Villaverde, J., Gillingham, K. T. & Scheidegger, S. Climate change through the lens of macroeconomic modeling. Annu. Rev. Econ. 17, 125–150 (2025).Article Google Scholar Mandel, A., Battiston, S. & Monasterolo, I. Mapping global financial risks under climate change. Nat. Clim. Change 15, 329–334 (2025).Article ADS Google Scholar Emmerling, J. et al. A multi-model assessment of inequality and climate change. Nat. Clim. Change 14, 1254–1260 (2024).Article ADS Google Scholar Orlov, A. et al. A better integration of health and economic impact assessments of climate change. Environ. Res. Lett. 19, 031004 (2024).Article PubMed PubMed Central Google Scholar Download referencesAcknowledgementsJ.S. acknowledges funding from the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under Germany’s Excellence Strategy—EXC 2037 “Climate, Climatic Change, and Society” (CLICCS)—Project Number: 390683824. We thank three anonymous referees for their constructive comments, Gunnell Sandanger and Miriam Stackpole Dahl for suggestions that improved the readability of the commentary, and Maura Dewey and Srinath Krishnan for valuable feedback.Author informationAuthors and AffiliationsCICERO Center for International Climate Research, Oslo, NorwayAnton Orlov & Jana SillmannResearch Unit for Sustainability and Climate Risks, Department of Earth System Sciences, University of Hamburg, Hamburg, GermanyJana SillmannAuthorsAnton OrlovView author publicationsSearch author on:PubMed Google ScholarJana SillmannView author publicationsSearch author on:PubMed Google ScholarContributionsA.O. developed the concept and wrote the manuscript. J.S. co-developed the concept and contributed to the manuscript writing and editing.Corresponding authorCorrespondence to Jana Sillmann.Ethics declarationsCompeting interestsThe authors declare no competing interests.Peer reviewPeer review informationNature Communications thanks Valeria Costantini, who co-reviewed with Mariagrazia D’Angeli; Christian L. E. Franzke and the other, anonymous, reviewer(s) for their contribution to the peer review of this work.Additional informationPublisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.Rights and permissionsOpen Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.Reprints and permissionsAbout this articleDownload PDF