MainDeforestation is a major contributor to global greenhouse gas emissions, accounting for at least 12% of total emissions worldwide1. In Brazil, one of the world’s largest sources of land-use change emissions2,3, this challenge is particularly acute because it is home to the Amazon rainforest, one of the planet’s largest and most important carbon sinks4. Growing evidence suggests that continued deforestation and forest degradation are pushing the Amazon towards a critical tipping point, beyond which areas could shift from being a net carbon sink to a net carbon source5. Such a transition would undermine Brazil’s climate mitigation efforts6 and have profound implications for the global climate system.The drivers of deforestation in Brazil are complex. Local governance factors such as weak enforcement of environmental regulations, insecure land tenure and illegal land grabbing remain central7,8,9. These are compounded by demand for agricultural commodities, especially for beef and soy10, as well as expanding infrastructure and market access11,12. Much of literature has focused on drivers that are proximate to the forest frontier, framing deforestation largely as a problem of poor enforcement or trade-offs between growth-oriented infrastructure and environmental sustainability.However, a recent World Bank study by Ferreira Filho and Hanusch13 introduces an alternative and potentially powerful perspective: deforestation is also shaped by the macro-structural characteristics of the Brazilian economy. Using a dynamic computable general equilibrium (CGE) model with spatially explicit land-use transitions, they show that that productivity-enhancing structural transformation towards more urban, less land-intensive traded sectors (such as manufacturing) can substantially reduce deforestation. These productivity gains weaken agriculture’s relative competitiveness, lowering the economic incentive to convert forest land.While Ferreira Filho and Hanusch’s model offers valuable insights into land use and emissions, it gives limited guidance into what such a transformation would mean for Brazilian workers. Yet, structural transformation is not just about shifts in productivity—it is also about shifts and reallocations for people. In practice, the feasibility and fairness of deforestation-reducing growth strategies depend heavily on employment outcomes, particularly for low-income, rural populations. As CGE models tend to abstract away from labour market frictions, such as skills mismatch, spatial mobility and sectoral switching costs14, they give little insight into the challenges workers face when adapting to structural change and risk overlooking critical barriers to a socially equitable transition15.To address this gap, we extend the macro-structural lens developed by Ferreira Filho and Hanusch13 by developing an agent-based labour market model that explicitly simulates how workers in different occupations and spatial locations are affected by productivity-led development scenarios. While we restrict this analysis to the scenarios analysed in Ferreira Filho and Hanusch13, our model is generalizable and could be leveraged to explore other scenarios and development trajectories in Brazil. Our approach builds on modelling, sustainability and economic network literatures; see ‘Related literature’ in the Supplementary Information for details.To capture labour market adjustment dynamics, we first construct a labour mobility network on the basis of historical job transition data in the Brazilian economy. This network reveals that skill and geographic mismatches significantly influence the ability of workers to shift into new employment. Workers are 62% more likely to transition into an occupation within the same broad category than into a different one and 7.2 times more likely to remain within their home state than relocate, highlighting both occupational and spatial rigidity.We extend the agent-based labour market model developed by del Rio-Chanona et al.16 to incorporate both spatial and skill-related barriers to mobility, drawing on our empirically derived labour mobility network. This allows us to simulate how different types of workers are likely to respond to productivity shocks across sectors and regions in Brazil. To avoid confusion with the five macroregions of Brazil, we use ‘region’ here to refer to the 16 regions of Brazil used in our study, following Ferreira Filho and Hanusch13. The regions are either states, for example, Bahia, or groups of states, for example, ‘RSudeste’, which is made up of Minas Gerais, Espírito Santo and Rio de Janeiro. We calibrate the model using quarterly unemployment and Gross Domestic Product (GDP) data and validate it through both in-sample and out-of-sample testing, demonstrating strong alignment with observed macro-economic trends. To further assess its robustness, we simulate the effects of the coronavirus disease 2019 (COVID-19) pandemic, an exogenous shock with well-documented labour market consequences, and show that the model successfully reproduces the sharp rise in unemployment experienced during this period.We apply the model to the scenarios developed by Ferreira Filho and Hanusch13 by exploiting relative changes in labour demand across sectors in relation to the baseline scenario. This reallocation generates substantial variation in occupation- and region-specific labour requirements, generating mobility frictions that our disaggregated agent-based model (ABM) is designed to capture. We preserve the CGE assumption that total labour demand in the scenarios is the same as in the baseline. However, as our model operates at the region–occupation level, layoffs and vacancies emerge even when aggregate employment remains unchanged. Workers respond by searching for jobs across sectors and locations, guided by empirically observed transition patterns. Using granular time steps, the model tracks the evolving number of employed and unemployed workers, as well as unfilled vacancies in each region and occupation.Our results show that without targeted policy support, nearly all productivity-driven scenarios worsen labour market inequality. High-wage and high-skill workers tend to experience lower unemployment rates compared with the baseline, while lower-wage workers—particularly those in agriculture—are more likely to face increased unemployment. Agricultural workers, who are among the least spatially and occupationally mobile, are negatively affected in almost every scenario.The scenario that delivers the best environmental outcomes, which involves a shift towards more productive manufacturing and reduced deforestation, improves employment prospects for mid- and high-skill workers. However, it also results in higher unemployment for agricultural and other low-skill workers. The only scenario in which agricultural workers are not adversely affected is one where agricultural productivity in the Amazon increases disproportionately. Here, however, deforestation and emissions rise significantly.To assess the impact of mobility-enhancing policies, we rerun the same set of scenarios with and without a policy intervention that increases agricultural workers’ mobility into occupations with similar skill levels but in other sectors. In the scenario with increased productivity in the manufacturing sector, this policy improves unemployment outcomes of Amazon agricultural workers by 50% on average. These results highlight how well-designed labour market policies can help align productivity gains with equitable employment outcomes, supporting more just and feasible economic development pathways in Brazil.ResultsOccupational and regional labour mobility frictionsOur model is informed by empirical worker mobility networks, which we generate using the Relação Anual de Informações Sociais (RAIS) dataset for the period 2011–2019. The RAIS dataset details all active employer–employee contracts in Brazil’s formal labour market (Methods). Figure 1a shows Brazil’s occupational mobility network, where each node represents an occupation, which can be linked to others on the basis of historical transitions by workers between them. Nodes belonging to the same broad occupational category cluster tightly together, indicating that transitions across occupational categories are relatively rare. This pattern suggests that occupational mobility is constrained by skill-related barriers, limiting workers’ ability to reallocate across sectors and broad occupational categories in response to changes in labour demand.Fig. 1: Labour mobility networks for Brazil.Full size imagea, The national occupational mobility network for Brazil, coloured by occupation group. Nodes represent occupations and links represent historical job transitions from 2011 to 2019. Number of observations = 2,591 occupations at the six-digit level. b, The regional occupational mobility network, coloured by region. Nodes represent occupation–region pairs and edges represent the frequency of transitions made between them. Number of observations = 6,496 occupation–region pairs. c, A map of Brazil, with states coloured similarly to in b. Administrative boundaries data in c from Instituto Brasileiro de Geografia e Estatística44. Model regions: Matopiba includes Tocantins, Maranhão and Piauí; RNordeste includes Ceará, Rio Grande do Norte, Paraíba, Pernambuco, Alagoas and Sergipe; RSudeste includes Minas Gerais, Espírito Santo and Rio de Janeiro; RSul includes Santa Catarina and Rio Grande do Sul; GoiasDF includes Goiás and Distrito Federal.Disaggregating labour flows further, Fig. 1b shows a regional occupational mobility network where nodes now represent occupation–region pairs linked to each other on the basis of past job transitions across both occupations and regions. Strikingly, the regional occupational mobility network closely resembles the geographical map of Brazil (Fig. 1c), highlighting the strong degree to which geography (and associated factors such as moving costs and familial ties) influence labour flows.We quantify these patterns using assortativity measures, which indicate how often occupation–region pairs connect to other pairs with the same occupation or region (Methods). Regional assortativity in the regional occupational network is high (0.77), while occupational assortativity at the one-digit level is also substantial (0.56), confirming that both spatial and occupational frictions play a central role in shaping labour mobility.Agent-based labour market modelThe regional occupational mobility network outlined above is a key component of the agent-based labour market model, which runs in monthly time steps and simulates how workers respond to changing labour demand across occupations and regions. Workers can be either employed or unemployed in an occupation–region pair (nodes of the network). Each month, employed workers may lose their jobs through two mechanisms: a spontaneous process representing routine turnover (at rate δu) or a state-dependent process driven by mismatches between realized and target labour demand (determined by the external scenarios, explained below). Similarly, vacancies open through both spontaneous creation (at rate δv) and state-dependent adjustments when target labour demand exceeds current employment. Unemployed workers search for new positions by applying to vacancies in occupations and regions connected to their previous job through the empirical mobility network, accounting for realistic labour market frictions. The matching process uses an urn-ball matching function: if at least one worker applies to a vacancy, one applicant is drawn uniformly at random and hired, capturing the coordination failures inherent in decentralized job markets (see ‘Search and matching process’ in the Supplementary Information for details). The model outputs time series of employment, unemployment and vacancies for each occupation–region pair.A key advantage of this model is its versatility for (soft) coupling to different macro-economic models. The model requires only a time series of labour demand at the occupation level, which informs target labour demand. When target demand exceeds current employment, new vacancies open; when realized demand exceeds target, workers are separated. The speed of these adjustments is controlled by parameter γ. We integrate the model with the CGE scenarios through one-directional coupling: labour demand projections inform how many workers each occupation–region should employ, but labour market outcomes do not feed back into macro-economic projections. This approach imposes realistic mobility frictions on labour demand scenarios while maintaining computational feasibility. While soft coupling has methodological limitations—such as differences in equilibrium assumptions between CGE and ABM frameworks—it represents a pragmatic first step towards integrating labour market frictions into established CGE models for policymaking17.We calibrate the three parameters (δu, δv and γ) to the Brazil economy using quarterly unemployment and nominal GDP data from the years 2012–2019 and validate it using out-of-sample data from 2020–2025. The model achieves strong in-sample performance with unemployment rate correlations exceeding 0.93 and successfully captures the COVID-19 unemployment spike during the out-of-sample validation period (see ‘Calibration and validation methodology’ in the Supplementary Information for details).Development pathways for BrazilHaving validated the model’s ability to reproduce Brazil’s labour market dynamics, we apply it to analyse selected growth path scenarios from Ferreira Filho and Hanusch13, which are presented in Table 1. These scenarios simulate how Brazil’s economy and environment might evolve from 2018 to 2030 if investments and technological improvements increased productivity by 0.5 percentage points (p.p.) annually in manufacturing, agriculture, services or mining. Productivity gains in less land-intensive sectors, such as manufacturing, reduce deforestation and emissions, while agricultural productivity gains can increase emissions due to agriculture’s carbon intensity.Table 1 Key scenario features with an overview of labour market outcomesFull size tableImportantly, while the CGE model also projects how these productivity changes reshape labour demand across industries and regions, our labour model requires occupation–region-level demand time series. We therefore convert the CGE model’s industry-level labour demand at each time step to occupation-level demand by applying the observed occupational composition within each industry (see Methods for more details on scenarios and coupling).Table 1 presents our labour market results alongside each scenario’s environmental and economic outcomes, showing which occupation–region pairs face the largest changes in unemployment and unfilled vacancy rates relative to a baseline.Across all scenarios, we find that high-wage occupations tend to experience unemployment rates below baseline, while low-wage occupations face higher unemployment. Vacancy rates change the most in the northern areas and tend to be higher than in the baseline for mid-level and professional occupations. While sectoral reallocation creates significant variation in unemployment and vacancy rates at the disaggregated level, aggregate unemployment rates are virtually unaffected. This is mostly due to the fact that, by the design of the CGE, aggregate labour demand in the scenarios equals the baseline’s demand.Agricultural workers are the most negatively affected in nearly all scenarios and regions. The unemployment rate for these workers is the highest across all large-occupation groups, with increases between 0.13 and 2.11 p.p. (Tables A1 and A2 in the Supplementary Information). Unfilled vacancy rates for agricultural jobs decline in most cases. The sole exception occurs in the scenario where agricultural productivity increases in the Amazon only. Here, agricultural workers in Amazonas, Amapa, Acre, Rondonia, Roraima and Mato Grosso experience lower unemployment than the baseline, and unfilled vacancy rates for agricultural jobs increase slightly in these regions (Fig. A5a in the Supplementary Information). However, in this scenario, emissions increase owing to an increase in agricultural land expansion into forest land13. Thus, improved labour market outcomes for Amazon agricultural workers occur only under the scenario with the worst environmental performance.Skill and spatial mismatches in emission-reducing pathwayThe nationwide manufacturing growth path delivers the best environmental outcomes because a more competitive manufacturing sector—relying on Brazil’s low-carbon energy supply—would divert resources away from more emission-intense activities, in particular agriculture. This growth path also generates the largest demand reallocation and labour market frictions. In this section, we examine further the distributional impacts of the nationwide manufacturing growth path to identify workers who may face negative distributional consequences and require policy support. We first analyse patterns at the broad occupational classification (one-digit) level and then for specific occupation–region pairs at the detailed (four-digit) level. See Figures A4 and A5 in the Supplementary Information for results for other scenarios.Geographic patterns of unemployment vary by occupation type; Fig. 2a shows unemployment changes with occupations ordered by wage and skill level, and regions from south (left) to north (right). Agricultural workers face the steepest unemployment increases in the South region, while industrial, repair and service workers experience their worst outcomes in the north and northeast, Brazil’s poorest areas. Professionals of Arts and Sciences experience the largest declines in unemployment rates across scenarios, benefiting the most from sectorial productivity gains. Middle-level technicians, which are in the middle of the wage scale, show improved unemployment outcomes and increased unfilled vacancy rates, suggesting mid-level technical jobs as options for workers in low-skill occupations facing persistently higher unemployment, although firms face difficulties filling these positions.Fig. 2: Average changes in unemployment and vacancy rates under nationwide productivity increases in manufacturing.Full size imagea, The average percentage-point change in the 2018–2030 unemployment rate from the baseline, for each one-digit occupation and region. b, The average percentage-point change in the 2018–2030 unfilled vacancy rate from the baseline for each one-digit occupation and region. Macroregions and skill levels displayed for reference: S, south; SE, southeast; CW, central-west; N, north; NE, northeast. Skill level based on the International Standard Classification of Occupations (ISCO), Adamczyk et al.45 and Lewandowski et al.46. BRL, Brazilian real. N = 5,474 occupation–region pairs.Conversely, vacancy patterns, as shown in Fig. 2b, reveal different spatial constraints by skill level. Changes in unfilled vacancy rates tend to concentrate almost exclusively in the remote northern states of Amapá, Acre and Roraima, which face a spatial–skill mismatch: unfilled vacancy rates are higher than baseline for high- and middle-skill occupations but lower than baseline (or increase less) for lower-skill workers.At the fine-grained (four-digit) occupation level, we find that skill and spatial frictions, as captured by a workers’ position in the regional occupational network, alter unemployment and vacancy rates relative to a frictionless benchmark. To illustrate this, we compare our baseline results against a complete network scenario (Kn), where unemployed workers can apply equally to any job opening without facing skill or spatial barriers. Figure 3a,c shows this comparison: the blue squares represent outcomes under the complete network, while the coloured points show outcomes with realistic mobility frictions. In the frictionless case, unemployment and vacancy rates cluster tightly around zero regardless of demand changes. By contrast, once mobility frictions are introduced, occupations with identical demand change face substantially different outcomes depending on their position in the network. This dispersion demonstrates that labour mobility frictions prevent some workers from moving to where jobs are available.Fig. 3: Manufacturing growth path—Brazil: unemployment and unfilled vacancy outcomes.Full size imagea, The percentage-point change of the 2018–2030 average unemployment rate from the baseline against percentage demand change compared with baseline in 2030; each data point represents an occupation–region pair, shaped by region and sized proportional to employment in 2018. b, The regional percentage-point change of the average unemployment rate from the baseline. c, The percentage-point change of the 2018–2030 average vacancy rate from the baseline against percentage demand change compared with baseline in 2030; each data point represents an occupation–region pair, shaped by region and sized proportional to employment in 2018. d, The regional percentage-point change of the average vacancy rate from the baseline. To highlight the role of skills and spatial frictions, we show the results for the case with no worker frictions in blue squares. Here, we run the model with a complete network (Kn), that is, with each occupation–region pair fully connected to all other occupation–regions, using equal weights. Under this formulation, unemployed workers do not face any skill or spatial frictions and so can apply for a job in any occupation–region pair that is hiring, greatly reducing the frictions within the model. In this benchmark case, there is little variation in outcomes of occupation–regions with the same change in demand and little impact on the unemployment vacancy rates compared with the baseline overall. N = 5,474 occupation–region pairs. Administrative boundaries data in b and d from the United Nations Office for the Coordination of Humanitarian Affairs47 under a Creative Commons license CC BY 4.0.For example, among industrial workers (craft) in São Paulo (Fig. 3a), ‘garment sewing machine operators’ and ‘hand weaving, handmade shoe and leather goods and related craftsmen’ face similar demand decreases but divergent unemployment effects—the former suffers increased unemployment, while the latter remains nearly unaffected. ‘Garment sewing machine operators’ cannot mitigate their demand decrease because their neighbouring occupations also experience declining demand and worse outcomes overall. The opposite pattern characterizes occupations surrounding ‘hand weaving, handmade shoe and leather goods and related craftsmen’ in São Paulo. This pattern is also observed between these same two occupations in the model region RSul, which comprises the states of Santa Catarina and Rio Grande do Sul.Among agricultural workers in RSul, ‘general agricultural workers’ and ‘agricultural support workers’ face higher increases in unemployment rates than ‘farm supervisors’ (about 0.95 p.p. higher than baseline versus only 0.07 p.p higher) despite similar changes to labour demand of about −9%. This happens because the occupations neighbouring ‘farm supervisors’ have much better unemployment outcomes than the close neighbours of general and support workers overall. In both examples above, such heterogeneity indicates that effective policies must take into account workers’ specific skills and the viability of their outside options, rather than treating all industrial and agricultural workers uniformly.Firms face substantial difficulties filling vacancies for ‘middle-level technicians’ and ‘professionals of sciences and arts’, with unfilled vacancy rates for these groups rising above baseline levels (Fig. 3c,d). However, unfilled vacancy rates for ‘industrial workers’, both machine and crafts, are either unchanged or decreased.The following example illustrates these differences in outcomes of ‘professional of arts and sciences’ and ‘industrial workers’: ‘civil and allied engineers’ in Roraima experience 7% higher labour demand than baseline and face vacancy rate increases of approximately 0.50 p.p. In comparison, ‘computer numerical control (CNC) machinists’ in Roraima face a larger demand increase (8.8%), yet their vacancy rate remains nearly unchanged. Although both occupations have few neighbouring occupations from which workers typically transition, the critical difference lies in the outcomes of these neighbours. The few occupations that historically have switched into ‘civil and allied engineers’ experience favourable demand changes and lower unemployment rates themselves, so that fewer workers are willing to make the switch now. By contrast, occupations neighbouring ‘CNC machinists’ face declining demand and lower vacancy rates in their own occupations, prompting workers to switch into ‘CNC machinists’ positions. These results indicate that in some cases relying solely on workers from similar occupations to move into high demand occupations may prove insufficient and that training programmes that can close larger skills gaps might be needed.Spatial frictions make it harder for firms in remote and smaller labour markets to hire workers, leaving some industrial vacancies unfilled despite increased unemployment rates within the occupational group (Fig. 3c,d). ‘Electric, telephone and data communication lines and cables installers and repairers’ in Acre and in GoiasDF both experience approximately 13% demand increases, yet vacancies remain unfilled for longer only in Acre. These spatial barriers in more isolated states—Acre, Amapa, Rondonia, and Roraima—create dual challenges where firms cannot fill positions even as unemployment rises elsewhere in the same occupations.We assess whether a policy that improves agricultural workers’ mobility into mid-level, industrial and repair-and-maintenance occupations could improve their outcomes in the more environmentally friendly pathways. We find that this policy improves Amazon agricultural workers’ unemployment outcomes by approximately 50%, with scenario-led increases in unemployment rates between 0.02 and 0.11 p.p. lower in model regions Acre, Amapa, Amazonas, Matopiba, Para, Rondonia and Roraima (see Fig. A7 in the Supplementary Information for details). In the Supplementary Information (Fig. A6), we show additional results for manufacturing productivity increases concentrated in the Amazon region only. In those cases, emissions are still reduced but the effects remain largely localized to northern states.DiscussionUnderstanding the labour market impacts of development pathways, alongside their environmental impacts, is vital for designing effective policies and securing public support. While much attention has been given to green transitions in advanced economies, greater focus on pathways for lower income countries is critical for global climate action18. This study builds on the insight that Brazil’s deforestation challenge is not just a matter of forest governance but also reflects deeper structural and productivity patterns in the macro-economy. While Ferreira Filho and Hanusch13 showed that shifts towards more productive, less land-intensive sectors such as manufacturing can reduce deforestation by weakening agriculture’s relative competitiveness, their model did not account for how such transitions affect workers. Here, we have addressed this gap by introducing and validating a hybrid modelling approach to assess how productivity-driven development pathways in Brazil could influence employment outcomes across occupations and regions.Our model allows workers to find job opportunities in other occupations and relocate to other regions, given empirically observed skill and spatial constraints. In view of these frictions, we find that low-skill, low-wage occupations are consistently the most adversely affected, highlighting the need for targeted policy support. In nearly all scenarios, higher skill occupations benefit, while agricultural workers face increased unemployment. When we simulate a modest policy intervention that increases mobility for agricultural workers, unemployment among these workers in the Amazon falls by 50%. These findings add to the growing evidence that green transitions, if unmanaged, may exacerbate existing inequalities19 and that retraining, reskilling and relocation policies will be vital to protect vulnerable workers 20.While our application focuses on scenarios from Ferreira Filho and Hanusch13, the modelling framework is generalizable. It could be extended to explore other transition scenarios in Brazil, including digitalization16, and used to identify occupations that may be at risk from multiple shocks hitting the labour force at once.Our model represents one of the most granular labour market models available so far, but there are opportunities for further work. The current analysis relies on formal labour market data only, corresponding to about 67%21 of the total labour force. Although the dataset we use accurately reflects nationwide demographics such as gender and age22, the lack of informal transitions might be particularly consequential for occupations with high levels of informality (such as in the services, agricultural and industrial-craft occupational groups) which are under-represented in our data. Incorporating data on informal labour from sources such as the Brazil National Household Sample Survey (PNAD), could improve coverage. The ABM could be extended to include wage dynamics and labour force entry/exit. In addition, hard coupling of the labour market ABM with an external macro-economic model, such as in refs. 23,24, would enable labour mobility frictions, wage adjustments and productivity losses to feed back into consumption and production decisions, potentially affecting long-term outcomes. This would allow unfilled vacancies to translate into lower industrial output. While our validation demonstrates strong model performance at the aggregate level, future work would benefit from validation at the occupation–region level.Our findings for Brazil echo those of Bergant, Mano and Shibata25 for the USA: a green transition may not dramatically shift employment geographically but requires different skills. Effective policy must therefore be based on a nuanced understanding of both spatial and skill-related frictions20,26,27,28. To this end, we contribute a new model with the adequate granularity that provides important insights: workers at the bottom of the wage distribution could face the greatest risks, while skill shortages in high-wage occupations and remote regions could constrain transition speed. Thus, skills development policies must be implemented early to ensure that the green transition can be both fast and does not increase labour market inequality.MethodsWe combine a regional occupational mobility network with an agent-based labour market model to estimate the effect of skill and spatial mismatches on unemployment and unfilled vacancy rates for a set of scenarios for Brazil’s economy. Below, we explain how we estimate and analyse the regional occupational mobility network. We then describe the labour market model, how the scenario outputs are integrated into it, and define the key metrics that summarize our model results.Regional occupational mobility networkWe extend the work of Mealy, del Rio-Chanona and Farmer29 on occupational mobility networks by incorporating regional mobility. Each node in the regional occupational mobility network represents an occupation–region pair. Weighted edges between them are informed by empirically observed job transitions.We construct the regional occupational mobility network using data from RAIS from 2011 to 2019. For each year, the RAIS dataset contains data on the population of all active employer–employee contracts in Brazil’s formal labour market.The network comprises 410 occupations and 16 regions. These 410 occupations are a hybrid level of the Classificação Brasileira de Ocupações (Brazilian Classification of Occupations; CBO)-2002 classification. We start with the four-digit level of the CBO-2002 classification, which has 570 occupations. To have a fully connected component and a consistent list of occupations across all regions, we merge some of these occupations into hybrid two- and three-digit codes. This merging results in 410 unique occupations. For some visualizations, we group the occupations into nine one-digit occupational groups, such as ‘professionals of sciences and arts’.For the network visualizations, we use the Force Atlas 2 (ref. 30) algorithm to position the nodes. Nodes repel each other, while an edge acts as an attraction force. The layout of nodes is created from the network structure alone, with no additional information about the nodes.In the model, a link between two occupation–region pairs is the empirical probability of moving between them as observed in the data between 2011 to 2019. To compute the transition probabilities, we count the number of workers that transitioned from occupation i in region α to occupation j in region β between 2011 and 2019, given by Tiα,jβ. The probability of a worker transitioning from occupation–region iα to occupation–region jβ is defined as$${A}_{i\alpha ,j\beta }=\frac{{T}_{i\alpha ,j\beta }}{{\sum }_{i,\alpha }{T}_{i\alpha ,j\beta }}.$$(1)Using population data avoids sampling errors, but, as RAIS does not include data on informal labour markets, our estimates may be biased if formal and informal workers display systematically different mobility patterns. We discuss that future work should address this issue.Using this occupation–region mobility network also relies on the assumption that historical transitions between occupations and regions are informative of potential future transitions. This assumption is grounded in skill-relatedness theory31. Originally developed to explain firm diversification patterns, this framework has been successfully extended to labour flows32. However, this approach may both over- and underestimate actual mobility. We probably overestimate frictions because our model maintains wage rigidity—workers cannot trade wage cuts for employment continuity. Evidence shows displaced workers do accept wage losses to avoid unemployment, although reservation wage adjustment is sticky33,34. Our fixed-population assumption omits worker entry and exit; this is a relevant model limitation, as replacement flows could partly align with structural change, ease some frictions and should be considered in future modelling and labour market policy design.AssortativityAssortativity measures the tendency of nodes to be connected with other nodes of similar type. Assortativity of a network with adjacency matrix A where element Akl gives the edge weight between nodes k and l, and nodes belonging to different types (x, y, z, ...) was originally defined by Newman35 and later extended to weighted networks by Yuan, Yan and Zhang36. We use weighted categorical assortativity, which is defined by Bücker et al.37 as equation 2 from Newman for assortativity with the fraction of edge weights rather than the fraction of edges. That is,$$\begin{array}{c}r=\displaystyle\frac{\sum _{x}({e}_{xx}-{a}_{x}{b}_{x})}{1-\sum _{x}{a}_{x}{b}_{x}}\end{array},$$(2)where exy is the fraction of all edge weights that join nodes with value x to nodes with value y, and ax and bx are the fraction of edges that start and end, respectively at nodes with value x. That is,$${e}_{xy}=\frac{{\sum }_{k\in x,l\in y}{A}_{kl}}{{\sum }_{k,l}{A}_{kl}},\qquad {a}_{x}=\sum _{y}{e}_{xy},\qquad {b}_{x}=\sum _{y}{e}_{yx}.$$(3)Labour market modelWe extend the agent-based labour market model developed by del Rio-Chanona et al.16 to include geographical mobility. Using the regional occupational mobility network, we model worker mobility between occupation–region pairs.We track the number of workers employed and unemployed, and open vacancies, at each occupation–region pair. At time t, let eiα;t be the number of workers employed in occupation–region pair iα, uiα;t the number of workers unemployed in occupation–region iα, and viα;t the number of vacancies open in occupation–region iα. Unemployed workers are assigned to the occupation–region pair where they were most recently employed.Workers quit or lose their job and vacancies are opened by two processes: a spontaneous process and a state-dependent process. The spontaneous process separates workers and opens vacancies at constant rates δu and δv, respectively. In the state-dependent process, workers are separated or vacancies are opened depending on the difference between the realized demand, diα;t = eiα;t + viα;t, and the target demand, \({d}_{i\alpha ;t}^{\dagger }\) at rate γ. The target demand takes the form of labour demand for each occupation–region pair at each time step and is where the link to the external macro model is introduced. Unemployed workers apply for jobs in occupation–regions connected to their previous occupation–region in the mobility network, with application probabilities proportional to edge weights.Building on ref. 16, the dynamics of the mean employment, unemployment and vacancies can be approximated by the following equations:$${\bar{e}}_{i\alpha ,t+1}={\bar{e}}_{i\alpha ,t}-\underbrace{\left({\delta }_{u}{\bar{e}}_{i\alpha ,t}+\gamma (1-{\delta }_{u})\max \left\{0,{\overline{d}}_{i\alpha ,t}-{d}_{i\alpha ,t}^{\dagger }\right\}\right)}_{{\text{separated}}\;{\text{workers}}}+\mathop{\underbrace{\sum _{j,\beta }{\overline{f}}_{j\beta,i\alpha,t+1}}}\limits_{\mathrm{hired}\;\mathrm{workers}},$$(4)$${\bar{u}}_{i\alpha ,t+1}={\bar{u}}_{i\alpha ,t}+\underbrace{\left({\delta }_{u}{\bar{e}}_{i\alpha ,t}+\gamma (1-{\delta }_{u})\max \left\{0,{\overline{d}}_{i\alpha ,t}-{d}_{i\alpha ,t}^{\dagger }\right\}\right)}_{{\text{separated}}\;{\text{workers}}}-\mathop{\underbrace{\sum _{j,\beta }{\overline{f}}_{i\alpha, j\beta,t+1}}}\limits_{\text{job-finding}\;\mathrm{workers}},$$(5)$${\bar{v}}_{i\alpha ,t+1}={\bar{v}}_{i\alpha ,t}+\mathop{\underbrace{\left({\delta }_{v}{\bar{e}}_{i\alpha ,t}+\gamma (1-{\delta }_{v})\max \left\{0,d^\dagger_{i\alpha ,t}-{\overline{d}_{i\alpha ,t}}\right\}\right)}}\limits_{\rm{opened\;vacancies}}-\mathop{{\underbrace{\sum _{j,\beta }{\bar{f}}_{j\beta,i\alpha,t+1}}}}\limits_{{\rm{hired}}\,{\rm{workers}}},$$(6)where \({\bar{f}}_{i\alpha ,j\beta ,t+1}\) represents the flow of workers from occupation–region iα to jβ. Each equation follows a stock-flow structure: employment in the previous period, minus separations (both spontaneous at rate δu and state dependent via the γ term when realized demand exceeds target), plus hiring flows. The unemployment equation adds separated workers and subtracts those finding new jobs, while the vacancy equation adds new openings (both spontaneous at rate δv and state dependent when target exceeds realized demand) and subtracts filled positions (see ‘Model derivation’ in the Supplementary Information for all modelling details).These equations enable efficient model iteration and calibration with Brazilian data. Following suggestions from ref. 38 for calibrating data-driven ABMs, we use time series data for calibration (unemployment rate and nominal GDP, 2012–2019) and validation (2020–2024). As Brazilian vacancy data are unavailable, we estimate vacancy rate ranges using interpolation from comparable economies. We demonstrate robustness across the calibrated parameter range (see ‘Calibration and validation methodology’ in the Supplementary Information for calibration details).Transition scenarios and model integrationWe study how skill and spatial frictions impact key labour market outcomes in productivity-driven scenarios for Brazil’s economy and environment developed by Ferreira Filho and Hanusch13. The authors use a CGE model that includes a land-use module to simulate growth, deforestation and emissions outcomes in Brazil during the 2018–2030 period. This setup reflects Brazil’s economic and emissions profile: the country is among the largest greenhouse gas emitters3, primarily due to emissions related to deforestation and agriculture39. However, agriculture is a highly productive sector in Brazil’s economy. Brazil aims, as set out in its Nationally Determined Contribution (NDC)6, to develop sustainably while lowering emissions by 53% in 2030 compared with 2005 levels and reach climate neutrality by 2050. Ferreira Filho and Hanusch13 try to understand how different sectoral productivity pathways relate to these decarbonization targets.The authors developed ten scenarios, each driven by productivity growth in different sectors and regions. We focus on the scenarios of single-sector productivity gains (that is, manufacturing, agriculture, services or mining), including both those with productivity shocks that affect the whole country and those affecting the Amazon region only (we cover eight out of the ten scenarios available as a result). We discuss the key findings that are common across all scenarios and deep-dive into the scenario that delivers the best environmental outcomes yet causes the largest labour market frictions, namely, the manufacturing growth path—Brazil. Results for all eight scenarios and additional discussions are presented in ‘Supplementary results’ in the Supplementary Information.The model timeline in Ferreira Filho and Hanusch13 is as follows: the authors build a baseline from 2015 to 2030. The 2015–2018 period is the historical period during which the model reproduces the observed patterns of Brazil’s economy. The scenarios for productivity increases start in 2019 and all results are interpretable in relation to the baseline.The scenarios output annual labour demand per region and 38 industries, while our model works with occupational demand. Therefore, to couple the CGE with the ABM labour market model, for each scenario and for the baseline, we convert sector-level labour demand to occupation-level labour demand. We map the CGE industries to the official industry classification in Brazil used in RAIS (National Classification of Economic Activities (CNAE)). We then use the average occupation employment per industry in 2018 and 2019 to convert industry-level labour demand to occupation-level labour demand per region. We assume the occupational breakdown per industry is stable over the study period. This is a simplifying assumption that we accept given our relatively short-term extrapolation to 203040 (but may lose accuracy over time with increased automation and technical change41).Second, the CGE model includes an assumption on population growth, but the labour market ABM can only handle a fixed population size. Therefore, we normalize all eight scenarios and the baseline with the total employment in the baseline scenario. If total labour demand for occupation–region pair iα in year τ in scenario s is Diα,τ,s, then the adjusted target demand for labour, \({D}_{i\alpha ,\tau ,s}^{* }\), is$${D}_{i\alpha ,\tau ,s}^{* }=\frac{{D}_{i\alpha ,\tau ,s}{\sum }_{i,\alpha }{D}_{i\alpha ,2018,{\mathrm{baseline}}}}{{\sum }_{i,\alpha }{D}_{i\alpha ,\tau ,{\mathrm{baseline}}}}.$$(7)Finally, within each year, we interpolate the demand linearly to calculate the monthly time series of target labour demand (\({d}_{i\alpha ;t}^{\dagger }\)) that is input into the labour market model.We use a soft-coupling, that is, a unidirectional, approach to integrate the CGE scenarios with our labour market model. The CGE-derived labour demand scenarios inform the ABM labour market dynamics, but labour market outcomes do not feed back into the macro-economic model. This approach has limitations, as the effects of skill and spatial frictions do not feed back into production decisions, but it allows us to incorporate realistic labour mobility frictions into macro-economic analysis while maintaining computational feasibility and leveraging the policy credibility of established CGE frameworks.RatesThe unemployment rate of occupation–region pair iα represents the workers that lost their job while working in occupation–region pair iα and have subsequently not found new employment. The average unemployment rate over the study period for occupation–region pair iα is defined as$${u}_{i\alpha ,{\mathrm{average}}}(T)=\frac{{\sum }_{t\in T}{u}_{i\alpha ,t}}{{\sum }_{t\in T}\left({u}_{i\alpha ,t}+{e}_{i\alpha ,t}\right)},$$(8)where T is the set of time steps that correspond to the years of transition, 2018–2030. As our model is normalized so that the size of the labour force (given by uiα;t + eiα;t) remains constant over time, this formulation is equivalent to computing average annual unemployment rates.The average x-month vacancy rate \({v}_{i\alpha ;\mathrm{average}}^{\left(\ge x\right)}\) is the fraction of vacancies over total realized demand in occupation–region pair iα that have been open for at least x months, on average, over the study period:$${v}_{i\alpha ,{\mathrm{average}}}^{\left(\ge x\right)}(T)=\frac{{\sum }_{t\in T}{v}_{i\alpha ,t}^{\left(\ge x\right)}}{{\sum }_{t\in T}\left({v}_{i\alpha ,t}+{e}_{i\alpha ,t}\right)}.$$(9)Throughout this study, we use the average 6-month vacancy rate, but the results are qualitatively unaffected by choosing the duration of an unfilled vacancy to be 3, 6 or 12 months.Occupational versus regional network effectsWe use variance decomposition to understand the contribution of between-region versus between-occupation variation to the total variation of unemployment and vacancy rate outcomes. Based on ref. 42, we decompose the variance of the unemployment rate u as$$\underbrace{\frac{1}{N}{\sum }_{i,\alpha }{({u}_{i\alpha }-{\bar{u}})}^{2}}_{{\text{total}}\;{\text{variance}}}=\underbrace{\mathrm{var}({\bar{u}}_{\alpha })}_{\text{between-region}}+\underbrace{\mathrm{var}({\bar{u}}_{i})}_{\text{between-occupation}}+\underbrace{\mathrm{var}({u}_{i\alpha }-{\bar{u}}_{\alpha }-{\bar{u}}_{i})}_{{\text{residual}}\;{\text{variance}}},$$(10)where N is the total number of occupation–region pairs, \({\bar{u}}_{\alpha }\) is the mean unemployment rate change within region α, \({\bar{u}}_{i}\) is the mean unemployment rate change within occupation i, and \(\bar{u}\) is the overall mean unemployment rate change across all occupation–region pairs. The variance decomposition for vacancy rates follows analogously.Data availabilityLabour mobility data from RAIS (Annual Social Information Report) are confidential but can be requested from Brazil’s Ministry of Labour and Employment. Labour demand scenarios can be requested from the authors of ref. 13. See corresponding author at https://doi.org/10.1016/j.forpol.2026.103834. Quarterly nominal GDP data are available via Ipeadata at https://www.ipeadata.gov.br/Default.aspx. Quarterly employment and vacancy rates are available via Eurostat at https://ec.europa.eu/eurostat. Data on Brazil’s Vacancy Index are available at https://www.fipe.org.br/pt-br/indices/catho-fipe/#indicador-mensal.Code availabilityThe complete code is available via Zenodo at https://doi.org/10.5281/zenodo.21476234 (ref. 43).ReferencesVan der Werf, G. R. et al. CO2 emissions from forest loss. Nat. Geosci. 2, 737–738 (2009).Article CAS Google Scholar Brasika, I. B. M. et al. Uncertainties in carbon emissions from land use and land cover change in Indonesia. Biogeosciences 22, 3547–3561 (2025).Article CAS Google Scholar UNEP-CCC. Emissions Gap Report (UNEP, 2021).Trancoso, R. Changing Amazon deforestation patterns: urgent need to restore command and control policies and market interventions. Environ. Res. Lett. 16, 041004 (2021).Article Google Scholar Gatti, L. V. et al. Amazonia as a carbon source linked to deforestation and climate change. Nature 595, 388–393 (2021).Article CAS Google Scholar Brazil First NDC 2023 Adjustment https://unfccc.int/sites/default/files/NDC/2023-11/Brazil%20First%20NDC%202023%20adjustment.pdf (UNFCCC, 2023).Furumo, P. R. et al. Land conflicts from overlapping claims in Brazil’s rural environmental registry. Proc. Natl Acad. Sci. USA 121, e2407357121 (2024).Article CAS Google Scholar Moutinho, P. & Azevedo-Ramos, C. Untitled public forestlands threaten Amazon conservation. Nat. Commun.14, 1152 (2023).Ferrante, L., B. Andrade, M. & M. Fearnside, P. Land grabbing on Brazil’s Highway BR-319 as a spearhead for Amazonian deforestation. Land Use Policy 108, 105559 (2021).Article Google Scholar Ritchie, H. Drivers of deforestation. Our World in Data (2021).Araujo, R., Assunção, J. & Bragança, A. Transportation infrastructure and deforestation in the Amazon. J. Dev. Econ. 177, 103559 (2025).Article Google Scholar Cezne, E. & Otsuki, K. The Amazon’s road to growth? Infrastructural imaginaries of Brazil’s BR-319 Highway. Econ. Soc. 54, 620–646 (2025).Article Google Scholar Hanusch, M. & de Souza Ferreira Filho, J. B. Breaking Erysichthon’s curse: unlocking economic growth while sparing forests in Brazil. For. Policy Econ. 189, 103834 (2026).Article Google Scholar Mealy, P. A., Bucker, J. J. J. H., de Moura, F. S. & Knudsen, C. Beyond Green Jobs: Advancing Metrics and Modeling Approaches for a Changing Labor Market(The World Bank, 2025).Knudsen, C., de Moura, F. S., Bucker, J. J. J. H. & Mealy, P. Five Frictions: Key Labor Market Barriers to Unlocking Job Growth in the Green Transition (The World Bank, 2025).del Rio-Chanona, R. M., Mealy, P., Beguerisse-Díaz, M., Lafond, F. & Farmer, J. D. Occupational mobility and automation: a data-driven network model. J. R. Soc. Interface 18, 20200898 (2021).Article Google Scholar Filatova, T. et al. The power of bridging decision scales: model coupling for advanced climate policy analysis. Proc. Natl Acad. Sci. USA 122, e2411592122 (2025).Article CAS Google Scholar Caucheteux, J., Fankhauser, S. & Srivastav, S. Climate change mitigation policies for developing countries. Rev. Environ. Econ. Policy 19, 69–89 (2025).Article Google Scholar Sato, M. et al. Skills and Wage Gaps in the Low-carbon Transition: Comparing Job Vacancy Data (Grantham Institute, LSE, 2023).Vandeplas, A., Vanyolos, I., Vigani, M. & Vogel, L. The Possible Implications of the Green Transition for the EU Labour Market (European Commission, 2022).Ulyssea, G. Firms, informality, and development: theory and evidence from Brazil. Am. Econ. Rev. 108, 2015–47 (2018).Article Google Scholar De Negri, J. A., Furtado de Castro, P., Ribeiro de Souza, N. & Arbache, J. S. Mercado Formal de Trabalho: Comparação Entre os Microdados da RAIS e da PNAD (Instituto de Pesquisa Econômica Aplicada, 2001).Dosi, G., C. Pereira, M., Roventini, A. & Virgillito, M. E. The labour-augmented K+S model: a laboratory for the analysis of institutional and policy regimes. EconomiA 21, 160–184 (2020).Article Google Scholar Wiese, S. et al. Forecasting macroeconomic dynamics using a calibrated data-driven agent-based model. Preprint at https://doi.org/10.48550/arXiv.2409.18760 (2024).Bergant, K., Mano, R. C. & Shibata, I. From Polluting to Green Jobs: a Seamless Transition in the US? (International Monetary Fund, 2022).Lim, J., Aklin, M. & Frank, M. R. Location is a major barrier for transferring US fossil fuel employment to green jobs. Nat. Commun. 14, 5711 (2023).Kuhn, M., Manovskii, I. & Qiu, X. The geography of job creation and job destruction. J. Monet. Econ. 158, 103898 (2026).Article Google Scholar Farinha, M. T., Balland, P.-A., Morrison, A. & Boschma, R. What drives the geography of jobs in the US? Unpacking relatedness. Ind. Innov. 26, 988–1022 (2019).Article Google Scholar Mealy, P., del Rio-Chanona, R. M. & Farmer, J. D. What you do at work matters: new lenses on labour. SSRN https://doi.org/10.2139/ssrn.3143064 (2018).Jacomy, M., Venturini, T., Heymann, S. & Bastian, M. ForceAtlas2, a continuous graph layout algorithm for handy network visualization designed for the Gephi software. PLoS ONE 9, 1–12 (2014).Article Google Scholar Neffke, F. & Henning, M. Skill relatedness and firm diversification. Strateg. Manag. J. 34, 297–316 (2013).Article Google Scholar Neffke, F., Sbardella, A., Schetter, U. & Tacchella, A. Economic Complexity Analysis. Papers in Evolutionary Economic Geography (PEEG) 2430 (Utrecht University, Department of Human Geography and Spatial Planning, Group Economic Geography, 2024).Farber, H. S. Employment, hours, and earnings consequences of job loss: US evidence from the displaced workers survey. J. Labor Econ. 35, S235–S272 (2017).Article Google Scholar Krueger, A. B. & Mueller, A. I. A contribution to the empirics of reservation wages. Am. Econ. J. Econ. Policy 8, 142–179 (2016).Article Google Scholar Newman, M. E. J. Mixing patterns in networks. Phys. Rev. E 67, 026126 (2003).Article CAS Google Scholar Yuan, Y., Yan, J. & Zhang, P. Assortativity measures for weighted and directed networks. J. Complex Netw. 9, cnab017 (2021).Article Google Scholar Bücker, J., del Rio-Chanona, R. M., Pichler, A., Ives, M. C. & Farmer, J. D. Employment dynamics in a rapid decarbonization of the US power sector. Joule 9, 101803 (2025).Article Google Scholar Pangallo, M. & del Rio-Chanona, R. M. Data-driven economic agent-based models. In The Economy as an Evolving Complex System IV (eds del Rio-Chanona, R. M. et al.) 308–341 (SFI Press, 2026).Sistema de Estimativas de Emissões e Remoções de Gases de Efeito Estufa (SEEG, 2023).Montt, G., S. Wiebe, K., Harsdorff, M., Simas, M., Bonnet, A. & Wood, R. Does climate action destroy jobs? An assessment of the employment implications of the 2-degree goal. Int. Labour Rev. 157, 519–556 (2018).Article Google Scholar Atalay, E., Phongthiengtham, P., Sotelo, S. & Tannenbaum, D. The evolution of work in the United States. Am. Econ. J. Appl. Econ. 12, 1–34 (2020).Article Google Scholar Baumgarten, D., Felbermayr, G. & Lehwald, S. Dissecting between-plant and within-plant wage dispersion: evidence from Germany. Ind. Relat. 59, 85–122 (2020).Article Google Scholar Berryman, A. et al. Code for “Skill and spatial mismatches for sustainable development in Brazil”. Zenodo https://doi.org/10.5281/zenodo.21476234 (2026).Malhas Territoriais: Malha Municipal Digital 2025 (Instituto Brasileiro de Geografia e Estatística, 2025).Adamczyk, W., Ehrl, P. & Monasterio, L. Skills and employment transitions in Brazil. Int. Labour Rev. 163, 199–224 (2024).Article Google Scholar Lewandowski, P., Park, A. & Schotte, S. The Global Distribution of Routine and Non-routine Work (Instytut Badań Strukturalnych, 2020).Brazil Administrative Level 0–2 Shapefiles (United Nations Office for the Coordination of Humanitarian Affairs, 2020).Download referencesAcknowledgementsWe thank J. Bento de Souza Ferreira Filho from Escola Superior de Agricultura ‘Luiz de Queiroz’ (ESALQ), University of São Paulo (USP), for providing scenario data.FundingThis work was partially funded by the UK’s Department for Energy Security and Net Zero and the Children’s Investment Fund Foundation (CIFF), under the Economics of Energy Innovation and System Transition (EEIST) programme (www.eeist.co.uk). The contents of this paper represent the views of the authors and should not be taken to represent the views of the UK government, CIFF or the organizations to which the authors are affiliated, or of any of the sponsoring organizations. This study was financed in part by the Coordenação de Aperfeiçoamento de Pessoal de Nível Superior—Brasil (CAPES; grant no. 001). J.D.F. thanks B. Gifford for their generous support of his work.Author informationAuthor notesThese authors contributed equally: Anna K. Berryman, Fernanda Senra de Moura, R. Maria del Rio-Chanona.Authors and AffiliationsMathematical Institute, University of Oxford, Oxford, UKAnna K. BerrymanInstitute for New Economic Thinking at the Oxford Martin School, University of Oxford, Oxford, UKAnna K. Berryman, Fernanda Senra de Moura, Joris Bücker, Pete Barbrook-Johnson, Penny Mealy & J. Doyne FarmerThe World Bank, Washington, DC, USAFernanda Senra de Moura, Marek Hanusch & Penny MealyEnvironmental Change Institute, University of Oxford, Oxford, UKFernanda Senra de Moura & Pete Barbrook-JohnsonSmith School of Enterprise and the Environment, University of Oxford, Oxford, UKJoris Bücker & J. Doyne FarmerDepartment of Arts and Sciences, University College London, London, UKPete Barbrook-JohnsonMonash Business School, Monash University, Clayton, Victoria, AustraliaPenny MealySanta Fe Institute, Santa Fe, NM, USAPenny Mealy & J. Doyne FarmerMacrocosm, New York, UKJ. Doyne Farmer & R. Maria del Rio-ChanonaDepartment of Computer Science, University College London, London, UKR. Maria del Rio-ChanonaThe Bennett Institute for Public Policy, Cambridge, UKR. Maria del Rio-ChanonaComplexity Hub Vienna, Vienna, AustriaR. Maria del Rio-ChanonaAuthorsAnna K. BerrymanView author publicationsSearch author on:PubMed Google ScholarFernanda Senra de MouraView author publicationsSearch author on:PubMed Google ScholarJoris BückerView author publicationsSearch author on:PubMed Google ScholarPete Barbrook-JohnsonView author publicationsSearch author on:PubMed Google ScholarMarek HanuschView author publicationsSearch author on:PubMed Google ScholarPenny MealyView author publicationsSearch author on:PubMed Google ScholarJ. Doyne FarmerView author publicationsSearch author on:PubMed Google ScholarR. Maria del Rio-ChanonaView author publicationsSearch author on:PubMed Google ScholarContributionsA.K.B. and F.S.d.M. implemented the main analysis, with input and additional analysis from J.B. and guidance from M.d.R.-C. and P.M. M.d.R.-C. conceived the model. A.K.B. and M.d.R.-C. carried out the code development and calibration, with input from F.S.d.M. All authors contributed to the research design, interpretation of results, and writing and revision of the paper.Corresponding authorCorrespondence to R. Maria del Rio-Chanona.Ethics declarationsCompeting interestsThe authors declare no competing interests.Peer reviewPeer review informationNature Sustainability thanks Angelica Sbardella 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.Supplementary informationRights and permissionsOpen Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, 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 changes were made. 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/4.0/.Reprints and permissionsAbout this article