Integrated assessment models (IAMs) are important tools for shaping global climate policy. At the international level, they feature heavily in the IPCC reports (IPCC, 2023) and inform climate negotiations under the UNFCCC (Science in the UNFCCC negotiations, 2025). In the United States, they are used to calculate carbon prices for policy appraisal (Rennert et al., 2022; Stern et al., 2022). In the European Union, model-based evidence of macroeconomic impacts has become a near prerequisite for accepting new climate or energy policy proposals (Pollitt and Mercure, 2018). IAMs' widespread use and growing influence have naturally invited scrutiny from the research community, leading to the identification of several limitations. It has been suggested that widening the underlying methodological scope is needed to overcome these limitations (Donges et al., 2021; Keppo et al., 2021; Stern et al., 2022). The FRIDA v2.1 model represents one such effort to broaden the methodological foundations of integrated assessment. The purpose of this paper is to document the economy module built for FRIDA v2.1 and its contributions to addressing some of these limitations, which are described below.

Most IAMs rely on highly aggregated damage functions, translating the complex relationships between temperature increases and economic consequences into a direct reduction in GDP. These functions' empirical foundations are limited because of data constraints, making their parametrisation highly uncertain (Farmer et al., 2015; Pindyck, 2013, 2017). Model outcomes, including policy-relevant optimal emission trajectories and the social cost of carbon estimates, demonstrate extraordinary sensitivity to damage function specifications (Gillingham et al., 2018; Hänsel et al., 2020; Nordhaus, 2019). Furthermore, owing to their aggregate nature, these damage functions operate as statistical black boxes (Pindyck, 2017). This aggregation necessarily compromises the traceability of model assumptions, making it impossible to identify how climate damage disrupts specific macroeconomic systems and sectors. The resulting opacity in model structure constrains researchers' ability to test individual assumptions and validate specific climate damage mechanisms (Elster, 2015; Meadows and Wright, 2008).

Similarly, most IAMs feature heavily aggregated economy modules. Representations of the financial system, notably, are either absent or omit the complex dynamics of credit creation, financial intermediation, and monetary policy, despite substantial literature documenting the financial risks posed by climate change (Battiston et al., 2021a; Pollitt and Mercure, 2018). Researchers have described how climate change and resulting extreme weather events pose risks to the global financial system, ranging from physical damage to financial assets to transition risks from stranded assets (Battiston et al., 2017, 2021b; Lamperti et al., 2019; Mandel et al., 2025). The mounting evidence has made financial and monetary authorities increasingly attentive to climate-related risks (Carney, 2015; Kiley, 2021; NGFS, 2019) and even prompted amendments to financial stability policies (Brunetti et al., 2021; Giuzio et al., 2019), making the inclusion of finance in IAMs timely.

In addition to understanding the vulnerability of financial systems to climate impacts, researchers call for models that are able to explore how global finance is poised to mitigate climate change. Sanders et al. (2022) noted that IAMs without finance leave “a crucial gap” in our ability to conduct macro model-based analyses of climate policy. A disorderly transition – where delayed recognition of climate risk leads to sudden, tardy action – is probable and poses systemic risks. Such a green transition could beget financial instability (Battiston et al., 2017, 2021a; Carattini et al., 2023; Garcia-Jorcano and Sanchis-Marco, 2025; Ojea-Ferreiro et al., 2024). Both Sanders et al. (2022) and Stern et al. (2022) warned that a disorderly transition could involve large and rapid changes in the price of carbon, resulting in sudden changes in asset values, triggering system-wide financial distress. These volatile dynamics are outside the scope of most contemporary IAMs because of their equilibrium assumptions and omission of finance, which leads us to the next IAM limitation: their lack of business cycle dynamics and short-term phenomena with long-term consequences.

In the most comprehensive treatment of the topic to date, Annicchiarico et al. (2022) noted that integrating business cycles in climate policy analyses represents a frontier area where substantial research gaps remain. Empirical evidence supports the hypothesis that emissions are procyclical (Doda, 2014). This finding indicates that instruments such as a carbon tax may be more effective if adjusted to the business cycle (Annicchiarico et al., 2022). Most contemporary IAMs are ill suited for evaluating the dynamic relationship between business cycles and climate policy (Stern et al., 2022). Adequately modelling these dynamics requires the integration of short-term mechanisms within a modelling framework that has traditionally favoured a long-term perspective (Pollitt and Mercure, 2018).

Neglecting short-term volatile mechanisms has implications beyond policy assessment. For example, extreme climate events can affect both existing economic activities and planned investments (Griffin et al., 2019). This is borne out in the growing body of research documenting the financial impact of climate change, especially in regard to failure rates of investments, leading to increased bankruptcies (Bartsch et al., 2024; Carattini et al., 2023; Feng et al., 2024). These unexpected defaults produce short-term negative employment shocks and drive long-term changes in lending standards, as the financial sector adapts its future growth expectations accordingly, negatively affecting investment and therefore both quantitative growth and qualitative economic development (Dell'Ariccia and Marquez, 2006; Fishman et al., 2024; Lown and Morgan, 2006; Rodano et al., 2018).

A more realistic representation of financial and monetary mechanisms and their dynamics further supports the endogenous modelling of innovation. Stern (2016) emphasised that most IAMs fail to capture the feedback loops in innovation processes, particularly the interactions across the economy that drive institutional and behavioural change. Mercure et al. (2019) extended this critique, showing that outcomes in climate–economy models are vastly different depending on how innovation is represented. They stress that innovation cannot be treated as an exogenous cost-reducing trend, as this approach fails to take into account the disruptive nature of the innovation process and its multiple feedbacks (Antonelli, 2017; Schumpeter, 1934) but should instead be endogenised within a framework that considers financial conditions, policy interventions, and other institutional dynamics – another direct call for economic disaggregation.

To address the highlighted gaps, the economy module built for FRIDA v2.1 rests on the following conceptual foundations: it is built on Schumpeterian theory, it operates in disequilibrium, all growth components are endogenous, and it features a stock-flow consistent financial architecture. These characteristics, elaborated on in Sect. 2, allow FRIDA v2.1 to analyse the two-way interactions of climate change through a multitude of direct and explicit impacts documented in Sect. 3. This opens new possibilities for understanding climate–economy interactions and evaluating policy interventions.

Table C1 in Appendix C provides an overview of IAMs with similar aims, alongside the classical DICE model for reference. While other models address some of the research gaps described above, they typically concentrate on the economy in isolation. FRIDA v2.1 stands out by including a similar level of detail in the other domains of the human–earth system (Schoenberg et al., 2025b), and is one of only two models with fully endogenous innovation.

FRIDA v2.1, and its novel economic module documented in this paper, aims to contribute to the recent research stream addressing these limitations. One body of work particularly influential for the FRIDA project is the “coupled human and natural systems” (CHANS) modelling approach, which emphasises that human and natural systems are defined by feedbacks, thresholds, nonlinearities, time lags, and emergence and link flows of matter, energy, and information (Alberti et al., 2011; Kramer et al., 2017; Liu et al., 2007). FRIDA v2.1 takes on this approach by disaggregating climate damage functions into multiple categories (see Wells et al., 2025, for a detailed treatment of the biophysical processes behind climate losses and damage, as represented in FRIDA v2.1), breaking down aggregate climate impacts into tangible, observable processes that can be empirically parameterised – processes that an aggregate damage function necessarily obscures. FRIDA v2.1's economy module adopts a disaggregated process-based approach to socioeconomic modelling that integrates short-term monetary, financial and innovation dynamics in the analysis.

These design choices make the model uniquely suited to address several underexplored research areas. A first potential application is the analysis of the financial and monetary consequences of climate change, and the role that these factors play in determining the macroeconomic costs of environmental damages. A second application is analyzing the complex relationship between innovation and climate change, including how the former could mitigate the latter, and how the latter could incentivize or disrupt the former. A third application is policy analysis. FRIDA v2.1 can be used for the comparative analysis of alternative funding strategies for climate policies, including income taxation, wealth taxation, government debt, central bank monetization and their potential combinations. Policy timing in relation to business cycles can also be explored: whether, for instance, a carbon tax introduced during a recession has different economic and environmental consequences than one introduced during an expansion. Some research efforts in these directions have already been made. Callegari et al. (2026) have assessed the relative economic and environmental performance of various innovation policy mix including growth-oriented and environment-oriented policies. Putranti et al. (2026) have analyzed how insurance policies could mitigate the financial instability effects of climate change. Ongoing research work is dedicated to analyzing how coordination between monetary and fiscal authorities could affect the cost-effectiveness of climate policies.

This documentation paper is structured as follows. Section 2 expands on the module's key conceptual foundations by establishing the theoretical rationale underlying model design choices and its attending limitations. Section 3 provides a detailed description of the module's structure across seven interconnected submodules, complete with stock and flow diagrams. Section 4 describes the calibration methodology and our protocol for the treatment of uncertainty. Section 5 presents simulation results demonstrating the model's performance against historical data and projection confidence intervals that explicitly communicate the project's inherent uncertainty. Section 6 concludes the paper with a run-through of contributions and a discussion of future applications. The model code and data are freely available on Github. Instruction on how to access them is provided in the Code and Data Availability section.

The Economy module is one of seven modules in FRIDA and represents the global economy. All internal processes are simulated in nominal terms, which is consistent with the module's monetary framework (real GDP is calculated by adjusting for inflation in the GDP submodule, as described in Sect. 3.6). It receives inputs from the Climate, Land Use and Agriculture, Demographics, and Energy modules and provides outputs to the Demographics, Resources, Energy and Behavioural Change modules. A schematic representation of these interactions is provided in Fig. 1. While embedded in FRIDA, the Economy module can also function independently as a stand-alone macroeconomic model of growth under exogenous environmental constraints. To explain the high-level interactions between all the modules and how the Economy module influences the rest of the model, see Schoenberg et al. (2025b).

The Economy module includes 761 equations, of which 33 are state variables. It is calibrated against data sourced for the period 1980–2023, covering the key macroeconomic indicators GDP, investment, consumption, government expenditure, the government debt-to-GDP ratio, inflation, unemployment, and wages. These time series were sourced from the World Bank, IMF, OECD, WID, and ILO, ensuring consistency and reliability (see Appendix A for the specific payoff elements and dataset references).

The internal structure of the Economy module consists of the seven submodules listed in the Economy module box in Fig. 1, each representing distinct but highly interconnected processes. Before turning to the detailed documentation, we provide a brief description of the seven submodules and their roles in the module's feedback structure. The Circular Flow submodule simulates firms' production and households' consumption. The Finance submodule mediates private investment affecting the flow of funds in and out of the circular flow, employment levels and interest rates. The Innovation submodule, together with the Finance submodule, governs R&D orientation and exploratory lending, determining the productivity growth that drives long-term economic development while displacing existing investments and workers in the short term. The Government submodule sets fiscal and monetary policy, with tax revenue, transfers, public debt, and the policy rate responding to conditions elsewhere in the economy. The Employment submodule translates investment flows into labour demand, wage formation, and unemployment, feeding back into consumption and government transfers. The GDP and Inflation submodules aggregate these flows into nominal and real output and determine the price level, which in turn affects consumption, wages, and the policy rate. This section documents the processes represented within each of these submodules and how they interact in more detail. First, the Circular Flow submodule is described, followed by the Finance, Innovation, Government, and Employment submodules, and finally the GDP and Inflation submodules.

FRIDA is implemented in Stella Architect 3.8 and simulated over a time horizon spanning 1980–2150. The integration method used to simulate the model is fourth-order Runga-Kutta (RK4) with a timestep of 1/8th of a year. A multistep protocol was developed for estimating parameter value ranges and presenting results in such a way as to communicate the range of possible outcomes given the input uncertainty rather than a single outcome (Schoenberg et al., 2025b).

The first step of the protocol is calibration using Powell's BOBYQA algorithm (Powell, 2009). Parameter inputs whose values, or value uncertainty ranges, could not be sourced from the data or literature were given wide uncertainty ranges on the basis of the outcomes created at the most extreme potential values for the parameter. The calibration involved varying the inputs within their individual uncertainty ranges to minimise the squared error between historical data and simulated data for 16 payoff elements for the period 1980–2023. The temporal coverage of the historical data series varies. The labour share of GDP for instance, sourced from the ILO, begins in 2004, which accounts for the absence of observed data before that year in Fig. 6g and h. A table listing all the payoff elements, with reference to the corresponding datasets, is found in Appendix A. This first step produced the best fit single-term endogenous model behaviour (EMB) shown in Fig. 6 for a selection of 12 key performance indicators.

The single-run EMB run, as shown in Fig. 6, presents the model's behaviour over the 1980–2023 calibration period, with simulated trajectories whose general level and variability are broadly consistent with the observed data. Notable features of these results include the model's ability to produce business cycle fluctuations, which are visible in the changing slope of GDP (Fig. 6a and b) and are particularly prominent in the unemployment rate (Fig. 6k) and the labour share of GDP (Fig. 6h). These cycles arise from the interaction of innovation and finance described in Sect. 3.2 and 3.3, and should not be interpreted as replicating specific historical downturns. Their amplitude and frequency are broadly consistent with observed macroeconomic variability, but their phase does not align perfectly with particular historical economic events, since FRIDA does not include shocks exogenous to the processes described in Sect. 3. This explains the absence of effects of the 2020 pandemic, 2009 banking crisis and similar events in the model results. For the historical period, the model does not include discrete policy changes, and endogenous policy changes are continuous. These deviations are consistent with FRIDA's design philosophy: rather than aiming for perfect replication of historical events driven by exogenous shocks, the model generates endogenous oscillations through the feedback processes within and across economic sectors.

The second step of the protocol involves exploring the possibility space, arising from the parameter uncertainty ranges. This is done in two steps. In the first step, the parameters are constrained to their likely ranges. The likely range is defined as the interval within which a parameter can be varied, while the likelihood of the model's output remaining above the 1/1000th of the best fit likelihood obtained during calibration. These ranges are then symmetrised within the originally provided widest possible bounds. This approach ensured that samples were distributed evenly around the default parameter set while still covering the regions with the highest likelihood. The resulting parameter bounds are documented in Appendix B.

On the basis of these ranges, we conduct a 100 000-member global sensitivity analysis (Saltelli, 2008) using Sobol sequence sampling (Sobol' and Levitan, 1999). This technique produces a set of sample points aimed at efficiently sampling the high-dimensional parameter space of the model. This method is needed because the high number of parameters to be sampled makes more systematic sampling computationally infeasible. The model is then run for each sample point, i.e., a set of parameter values, to produce the ensemble presented as the result of the model. As a result, the ensemble runs are not weighted by their likelihood, yet they can still be understood probabilistically: the resulting bounds highlight the regions of the output space that the ensemble most frequently explored, providing a probability-based interpretation conditioned on our sampling approach. It should be noted, however, that these confidence bounds are wider than they would be if likelihood weighting had been applied. While the sample is not weighted by likelihood, the spread captures a wide range of plausible system behaviours. The ensemble median, along with the 67 % and 95 % confidence intervals, are reported in Sect. 5.

This protocol reflects the project's commitment to openly communicating uncertainty, and it also underscores FRIDA's nature as an exploratory model: the goal is not to provide point predictions but to test the robustness of insights across a meaningful uncertainty space. Presenting single-line projections would understate the range of potential futures and could mislead stakeholders about the model's precision.

This section evaluates the model behaviour of the FRIDA v2.1 Economy module over a simulation time horizon spanning 2020–2150. The results are based on the 100 000-member ensemble run derived from the simulation protocol described in Sect. 4. Figure 7 shows the distribution of the ensemble members results for real GDP in selected years. As the runs progress the spread increases, exhibiting a fat tail. The ensemble median remains close to the calibrated best fit throughout the simulation period. This is a result of the uncertainty procedure varying the parameters symmetrically around the calibrated values. In the following figures the best fit run is not explicitly shown. It is treated as merely one of the ensemble members.

Figures 8 to 10 present the baseline results of the 100 000-member ensemble runs for key macroeconomic indicators. While Fig. 6 above presents FRIDA's behaviour over the historical calibration period, here we show the projected future economic trajectories under uncertainty. The figures show the medians and 67 % and 95 % confidence intervals of the 100 000 ensemble projections. While individual ensemble members have the oscillatory behaviour discussed above, the figures, especially Fig. 8, display relatively smooth trajectories as they show statistical measures of the asynchronous oscillatory behaviour across ensemble members. The smoothness is therefore a property of the ensemble statistics rather than of the model dynamics. Each member oscillates, but members differ in the period, amplitude, and phase of their cycles because they are generated under different parameterisations. When the median is taken across members at each point in time, peaks in some members coincide with troughs in others, flattening it. The same characteristic widens the confidence intervals, since the spread reflects both projection uncertainty and the different cycle phases across the ensemble.

Figure 8 shows GDP and its components in both nominal and real terms (constant USD 2021). In nominal terms, private consumption (Fig. 8b), private investment (Fig. 8c), government expenditure (Fig. 8d), and, hence, GDP (Fig. 8a) increase exponentially, which is consistent with historical patterns. Much of this growth, however, is due to inflation (Fig. 8a). Nonetheless, most ensemble members also exhibit continuing growth in real GDP (Fig. 8e) until the end of the simulation time horizon, albeit not exponentially. Private consumption (Fig. 8f) is the largest component of GDP and the main driver of the long-term upwards trend in GDP. The decline in the median growth rate around the end of the century is due to a decline in the growth of private consumption and investment (Fig. 8f and g). The investment slowdown originates in the Finance submodule (Sect. 3.2), where the rising STA increases the failure rate on new loans. The result is higher lending standards and risk premia, which together dampen private investment and income growth. This slowdown is partly compensated by an increase in government expenditures, including public investments (Fig. 8h). This is in turn driven by increases in government transfers, described below. However, there are ensemble members within the 95 % confidence interval that exhibit negative growth in real terms (nominal growth remains positive), reflecting ensemble members whose parameterisations specify the strongest climate damage consistent with historical data when combined with uncertainty in other areas of the economic system.

The inflation rate (Fig. 9a), starting from historical levels, declines over time. As a consequence, the central bank safe interest (Fig. 9b) rate is reduced (see Sect. 3.4.2). Without quantitative easing measures (not currently implemented in FRIDA), the central bank is not able to maintain the inflation rate at the target of 2 %, with the safe interest rate near the zero lower bound towards the end of the simulation time horizon. Interest rates paid by private investors (Fig. 9d) do not decline as strongly as the safe interest rate does because of the risk of defaults (Fig. 9c) perceived by banks, driving a risk premium between safe and risky interest rates (see Sect. 3.2). The default rate itself is climate-sensitive: rising STA increase defaults on new loans, and the resulting defaults sustain the risk premium even as the safe interest rate falls. The widening gap between the safe and risky rates is therefore a direct channel through which climate impacts are transmitted to the cost of investment finance.

A share of private investment (Fig. 8c) flows into research and development with the potential to increase productivity. Figure 10a shows the effect of these exploratory loans on changing economy-wide productivity (see Sect. 3.3). Increases in productivity increase GDP but also contribute to redundancies among workers, contributing to an increasing unemployment rate. The unemployment rate is further affected by demographic contraction around the turn of the century, combined with downwardly sticky wages keeping labour costs relatively high and decreasing the investment growth rate, leading to fewer new job vacancies because of proximately default-driven, but ultimately climate-driven, economic slowdown (see Sect. 3.2.3). Furthermore, economic volatility causing defaults translates to increased firing due to missed profits, in addition to productivity-driven displacement. These factors combine to reduce the economy's capacity to finance ever-growing numbers of workers, as clearly reflected in the growing unemployment rate (Fig. 10b; see Sect. 3.5). The increasing unemployment together with demographic change (see Schoenberg et al., 2025b) drives an increase in government expenditure for transfer payments (Fig. 10d). Transfers are designed to be countercyclical in the model: the share of expenditure directed to transfers adjusts with the number of unemployed and with the minors and retiree cohorts, so a rising unemployment rate and a rapidly ageing population both enlarge public spending. Increasing transfer payments contribute to the government debt-to-GDP ratio (Fig. 9c). Increasing government debt increases government interest rates, further straining the budget, which reduces the government's ability to invest and pay for transfers and other government services.

Taken together, these results show that the FRIDA v2.1 Economy module produces internally consistent long-run macroeconomic dynamics under uncertainty. The widening confidence intervals for the results reflect the substantial uncertainty inherent in long-run economic projection under climate constraints. However, robust patterns do emerge: median nominal aggregates continue to rise, and real activity generally increases but slows towards the end of the century because financial, demographic, and labour-market constraints are sensitive to climate impacts. While median projections suggest a generally growing but decelerating economy, the ensemble intervals expose substantial risk arising from climate impacts on the financial system, which can generate persistent unemployment, increase transfer burdens and public debt, and halt growth.

A substantial body of research has identified critical limitations in contemporary IAMs: aggregate damage functions that operate as statistical black boxes, the absence of financial system representation despite mounting evidence of climate-related financial risks, a lack of business cycle dynamics and short-term mechanisms with long-term consequences, and exogenous treatment of innovation. The novel economy module of FRIDA v2.1 documented in this paper addresses these limitations. The module is defined by its institutionally disaggregated structure, which allows for the traceability of specific climate impact and policy pathways through distinct economic sectors and processes. Moreover, it enables the model to leverage more observable economic variables with available time series data, strengthening its empirical foundations beyond the highly uncertain parametrisations of aggregate damage functions.

While FRIDA v2.1 addresses several critical gaps in the IAM literature, it introduces others that are inherent to its approach. Its global aggregation masks local climate impacts, precluding geographically differentiated policy analysis. The disequilibrium framework presents two distinct challenges. First, historical calibration necessarily attempts to reproduce observed fluctuations – including exogenous shocks such as pandemics and wars outside the model's scope – using only endogenous mechanisms, potentially overfitting to historical volatility. The resulting calibration should therefore not be interpreted as a replication of specific historical episodes, and the endogenous oscillations the model generates do not, and are not intended to, align perfectly in phase with observed business cycles. Second, the absence of utility maximisation assumptions limits the identification of optimal policy sets. Our ensemble methodology partially addresses these limitations by presenting results as probability distributions rather than point forecasts, explicitly foregrounding the considerable uncertainties inherent to long-run economic projection.

The model's primary contributions lie in its endogenous treatment of economic processes and the inclusion of finance, innovation, and business cycles. FRIDA v2.1 adopts a Schumpeterian disequilibrium framework that explicitly represents the role of the financial sector in both short- and long-term dynamics. This enables the analysis of climate-related financial risks, transition funding, and the potential for disorderly transitions to trigger system-wide instability – dynamics that remain largely unexplored in current policy-relevant models. The endogenous representation of innovation, driven by both incumbent R&D activities and the financing of new market entrants, captures the inherently disruptive nature of technological change and its dual role in both promoting productivity growth and displacing existing investments and workers.

The simulation results show that the model produces dynamics broadly consistent with observed macroeconomic patterns over the calibration period 1980–2023, and generates plausible projections through 2150. The ensemble approach, which is based on 100 000 simulation runs, produces widening confidence intervals over time, reflecting the compounding uncertainties as economic systems encroach on planetary boundaries. The median projections show that nominal aggregates increase throughout the simulation, whereas real activity expands more slowly and begins to decelerate as financial, demographic, and climate pressures accumulate. Moreover, the ensemble's more extreme scenarios reveal a substantial risk of economic volatility and even decline under more adverse climate impacts.

Following the historical period, the ensemble also reveals long-run constraints that arise from the model sectors' endogenous connections. Rising STA gradually weaken firms' ability to sustain investment by increasing failure rates and slowing growth. Moreover, demographic ageing and labour-market pressure increase the share of government spending devoted to transfers. Governments can rely on debt to maintain spending, but doing so increases interest burdens and inflationary pressure, further tightening fiscal conditions. These patterns – deteriorating investment conditions, rising welfare costs, and shrinking discretionary expenditures – emerge directly from the model's feedback structure rather than from imposed assumptions.

These endogenous constraints offer unique realism for policy experimentation with FRIDA v2.1. Climate-transition policy scenarios will operate within and interact with evolving financial, fiscal, and demographic pressures rather than being assessed in isolation. As a result, scenario outcomes reflect genuine trade-offs – for example, between maintaining welfare commitments, funding mitigation and adaptation, and preserving financial stability. By allowing these tensions to arise from the system's internal dynamics, FRIDA v2.1's economy module offers a rich exploration of transition pathways by embedding policy choices within the evolving macro conditions that will ultimately govern their success.