Modelling complex systems of heterogeneous agents to better design sustainability transitions policy
J.-F. Mercure, H. Pollitt, A. M. Bassi, J. E Viñuales, N. R. Edwards
MModelling complex systems of heterogeneous agents to better designsustainability transitions policy
J.-F. Mercure a,b, ∗ , H. Pollitt c , A. M. Bassi d , J. E. Vin˜uales b , N. R. Edwards e a Department of Environmental Science, Radboud University, PO Box 9010, 6500 GL Nijmegen, The Netherlands b Cambridge Centre for Environment, Energy and Natural Resource Governance (C-EENRG), University of Cambridge, 19 Silver Street,Cambridge, CB3 1EP, United Kingdom c Cambridge Econometrics Ltd, Covent Garden, Cambridge, CB1 2HT, UK d KnowlEdge Srl, 2, via San Giovanni Battista, 21057 Olgiate Olona (VA), Italy e Environment, Earth and Ecosystems, The Open University, Milton Keynes, UK
Abstract
This article proposes a fundamental methodological shift in the modelling of policy interventions for sustainabilitytransitions in order to account for complexity (e.g. self-reinforcing mechanisms, such as technology lock-ins, arisingfrom multi-agent interactions) and agent heterogeneity (e.g. di ff erences in consumer and investment behaviour arisingfrom income stratification). We first characterise the uncertainty faced by climate policy-makers and its implicationsfor investment decision-makers. We then identify five shortcomings in the equilibrium and optimisation-based ap-proaches most frequently used to inform sustainability policy: (i) their normative, optimisation-based nature, (ii) theirunrealistic reliance on the full-rationality of agents, (iii) their inability to account for mutual influences among agents(multi-agent interactions) and capture related self-reinforcing (positive feedback) processes, (iv) their inability to rep-resent multiple solutions and path-dependency, and (v) their inability to properly account for agent heterogeneity. Theaim of this article is to introduce an alternative modelling approach based on complexity dynamics and agent hetero-geneity, and explore its use in four key areas of sustainability policy, namely (1) technology adoption and di ff usion, (2)macroeconomic impacts of low-carbon policies, (3) interactions between the socio-economic system and the naturalenvironment, and (4) the anticipation of policy outcomes. The practical relevance of the proposed methodology issubsequently discussed by reference to four specific applications relating to each of the above areas: the di ff usion oftransport technology, the impact of low-carbon investment on income and employment, the management of cascadinguncertainties, and the cross-sectoral impact of biofuels policies. In conclusion, the article calls for a fundamentalmethodological shift aligning the modelling of the socio-economic system with that of the climatic system, for acombined and realistic understanding of the impact of sustainability policies. Keywords:
Environmental policy assessment, Climate change mitigation, Complexity sciences, Behavioural sciences
1. Introduction ff ect of uncertainty The starting-point of this article is the need to tackle the uncertainty facing climate policy-making and the relatedinvestment decision-making through a more realistic modelling approach.National and international public policy-making must confront the unprecedented challenge of e ff ectively manag-ing the complex interaction of economic development, energy systems and environmental change (IPCC, 2013b). Thee ff ects of stringent climate policies are subject to uncertainty and disagreement, which hinders policy action, and the ∗ Corresponding author: Jean-Franc¸ois Mercure
Email address:
[email protected] (J.-F. Mercure ) Tel: +
31 (0) 24 36 53254
Preprint submitted to Global Environmental Change February 10, 2016 a r X i v : . [ phy s i c s . s o c - ph ] F e b ack of policy clarity has, in turn, a chilling e ff ect on the private sector’s incentives to shift investment towards sustain-able options and opportunities. In contexts of damaging policy indecision, investors are often inclined to wait beforecommitting to new long-lived capital investment decisions (IEA, 2007). Meanwhile, carbon budgets are increasinglyconsumed (IPCC, 2013a), and the likelihood of avoiding dangerous climate change is rapidly decreasing. Consensuson desired outcomes achieved through international agreements (COP15 and COP21) urgently needs to be translatedinto consensus on actions, to identify e ff ective climate policy, and facilitate its rapid global adoption.At the roots of policy indecisiveness lie conflicts in our understanding of the complex interactions between tech-nology, society, the macroeconomy, and the environment. Policy-makers often consider that important trade-o ff s existbetween, on the one hand, improving the sustainability of the economy and, on the other hand, adequately supportingeconomic growth. The ensuing reluctance to act contrasts with the signals from an increasing body of reports pro-duced by think-tanks and international organisations arguing that wealth could be generated by new green technology.Decision-makers thus face dissonant signals, causing them to hold back their action. Indeed, successful public (andrelated private) initiatives of the scale required to stabilise emissions and adapt to climate change have no precedent,and they are unlikely to develop unless the important uncertainty as to their full implications is properly tackled.To make the analysis more intelligible, we identify four major areas where uncertainty contributes to climate policyindecisiveness: (1) the dynamics of technology adoption and di ff usion; (2) macroeconomic impacts of low-carbonpolicies; (3) interaction between human and environmental systems; and (4) policy implementation and e ff ectiveness.Solutions to environmental degradation rest on the di ff usion of mostly energy-related innovations, technologiesand practices throughout industries and between households. In many cases, low-carbon alternatives already exist.However, whether their adoption can be incentivised in time to avoid dangerous environmental change, and whetherthis is economically or technically possible, are open questions. Similarly, the extent to which such di ff usion couldsupport economic development is not well understood. Moreover, it is also unclear whether climate policies mayinfluence access to food, water and energy, and – if so – how. Hence, guidance on how to understand the complexinteractions between technology, the macroeconomy, and the environment is much needed.This article introduces a methodological approach that could significantly improve our ability to anticipate thee ff ects of climate policies, by integrating behavioural and non-equilibrium complexity science and environmentalfeedbacks into climate policy analysis, with a framework consistent across relevant disciplines. Equilibrium and optimisation-based approaches (e.g. Cost-Benefit Analysis or CBA, general equilibrium, cost-optimisation), despite their contribution to date (e.g. GEA, 2012, Stern, 2007), have five main shortcomings for theanalysis of the uncertainty identified in the preceding section.The first shortcoming concerns the concept underpinning these approaches, namely that of an e ff ective socialplanner coordinating society to minimise total cost or maximise aggregate utility, and the implicit corollary thatthere exists a unique stable economic equilibrium to which the economy tends to return after exogenous disturbances(e.g. see Mas-Colell et al., 1995). This method leads to approaches that are generally normative (i.e. they seek toidentify optimal strategies) rather than positive / descriptive (i.e. they do not always seek to describe actual systembehaviour with a high degree of realism). While a normative approach may appear attractive for policy purposes, itis fundamentally undermined by the fact that no such central coordination exists, while the assumption of optimalityneglects critical aspects of economic reality such as unemployment and market disequilbria that act both as drivers ofchange and opportunities for economic growth. If the economy is assumed to be permanently in an optimal state, thenplanning for and incentivising change makes little sense.A second, closely related shortcoming is that equilibrium theories do not su ffi ciently allow for the possibility thatagents may not be fully rational . Indeed, equilibrium theories typically involve finding the (inter-temporal) maximumof the aggregate utility function (or, in cost-optimisation models, the minima of total system costs). Such a maximum(minimum) results from the sum of the utility (cost) functions of the underlying agents, who are assumed to carryout an exhaustive ranking of their preferences over all possible products in all existing markets, and to optimise theirgoods basket choices given an economic context (see Mas-Colell et al., 1995). Underpinning this understanding is theassumption that the aggregate behaviour of a system of utility-maximisers can always be expressed as that of a singleaverage utility-maximising representative agent (see Kirman, 1992, for a critique). However, if one admits that agentsdo not carry out an exhaustive ranking of preferences ( bounded rationality ) or that agents may influence one another,2he utility of the representative agent becomes too complex or impossible to optimise due to increasing feedbacks andemerging complex dynamics. It then becomes unclear whether optimisation methods can be used at all, and whetherimposing strictly constant or decreasing returns, in order for models to converge, does not come at the price of losingtouch with reality. In other words, humans are not supercomputers optimising their choices over all goods o ff ered inall markets of the planet (Kirman, 1992). Agents usually know a small subset of information on goods they desire,and do not desire goods they know nothing of.Thirdly, equilibrium theories do not capture the possibility that agents may influence one another, leading topositive feedbacks and increasing returns . In this regard, the conventional equilibrium perspective may be termed reductionist , in the meaning ascribed to this term in complexity theory (Anderson, 1972), i.e. the macro system be-haviour is aggregated from micro properties, but without considering to a full extent the interactions (including mutualinfluence) between agents. Following complexity theory, one may consider interactions as additional elements thatlead to the emergence of additional collective phenomena . In economics, the inclusion of multi-agent interactions(specifically, the possibility that agents influence the behaviour of other agents) actually determines the di ff erencebetween, on the one hand, models of an economy e ff ectively formed of a single agent (or N isolated individuals) and,on the other hand, models of an economy where additional processes due to crowd e ff ects are allowed to emerge,including technology transitions and economic cycles. Allowing for multi-agent interactions is very important inpractice because such interactions are at the roots of al l self-reinforcing economic processes (crowd e ff ects), andthese are neglected in equilibrium economics. In the theory of complex systems, many properties emerge solely from the interactions between agents, not from the behaviour of the agents themselves (regardless of whether these arein homogeneous or heterogeneous contexts). This includes important economic phenomena such as the profile ofdi ff usion of innovations, learning-by-doing, expectations in finance and economic fluctuations, trends and fashions,technology lock-ins, and many more. These phenomena exist, but do not stem from microeconomic behaviour ofisolated individual agents. Therefore they cannot be studied with a methodological understanding that ignores inter-actions between agents. With multi-agent interactions, the representative agent may be understood to gain additionalemergent collective behavioural traits that the underlying agents do not possess when isolated. Thus while there is agood rationale for desiring a macro theory built on micro-foundations, the latter must include not just agent properties(e.g. preferences and income) but also inter-agent interactions, and this is a great – yet unavoidable – challenge.Because optimisation approaches (optimal growth, computable general equilibrium (CGE), partial equilibriumcost-optimisation) are sensitive to the curvature of their demand and supply or cost functions, as such, they are unfit tofully account for increasing returns , understood as self-reinforcing phenomena. This includes for example a declinein prices resulting from cumulative investment (learning), leading to investments that increase the likelihood of moresimilar investments. In an optimisation context, such positive feedbacks lead to numerical instability of the modelsolver (due to multiple solutions). Yet, processes with increasing returns are a very important feature of the realworld, particularly as regards climate policy. For example, early investments in solar energy may ultimately lead thetechnology, through learning, to competitiveness and possible market dominance, while a lack of early investmentswould have confined the technology to niche applications.A fourth, related shortcoming is the inability of conventional models to account for multiple solutions and path-dependence . Indeed, when increasing returns are introduced, several solutions to the optimisation problem emerge,and it becomes unclear which optimum is the correct one. In the investment example above, two di ff erent possiblefuture solutions evolve from di ff erent early investment decisions. Real-world technological and economic changeis thus path-dependent . Technology adoption typically follows S-shaped patterns, which stem partly from socialinfluence and interactions, where adoption of innovations increases the likelihood of further adoption of the sameinnovations (Rogers, 2010). Full path-dependency is a key property missing in many current economic models andIntegrated Assessment Models (IAMs), the latter combining global energy-economy-climate change phenomena usedto assess environmental policy, whether they are based on small (e.g. the DICE model Nordhaus 2013; FUND modelAntho ff and Tol 2014; PAGE model Hope 2011) or large datasets (e.g. PRIMES and GEM-E3 models, E3MLab 2013,2015; TIAM model, IEA / ETSAP 2012; MESSAGE model, IIASA 2013; AIM model, NIES 2012; REMIND model,PIK 2011).Finally, equilibrium theories do not su ffi ciently account for agent heterogeneity . People and firms are representedby the behaviour of a single representative agent with rational expectations. This agent is understood as the aggregatecollective behaviour emerging from the actions of the underlying agents, which can have distributed preferences (Mas-Colell et al., 1995). But no clear role is ascribed to di ff erences in the distribution of income or other socio-economic3nd industry parameters, and, following expected utility theory (EUT), only one type of behavioural response exists,namely a decision based on the expected utility associated with di ff erent choices times their respective probabilities(assumed to be known). In the real world, at least two types of deviations from EUT typically arise: behaviouraldiversity (variations around a central value, e.g. discrete choice theory, Domencich and McFadden, 1975) and be-havioural biases (systematic deviation from EUT, e.g. prospect theory, Kahneman and Tversky 1979; see also Sorrellet al. 2011 for a review and taxonomy). Agent heterogeneity can also be interpreted, in a utility maximisation per-spective, as some degree of ambiguity in terms of agent perceptions of optimality. Agent heterogeneity is importantin the representation of consumer or investor choices, which is critical in the process of the di ff usion of innovations,technologies and practices (Knobloch and Mercure, 2016). As can be inferred from standard innovation di ff usiontheory (Rogers, 2010), behavioural response, its diversity, the diversity of social groups, and unequal distribution ofinformation, are precisely what determines actual rates of adoption of innovations. This is standard knowledge in thefield of marketing research (e.g. Smith, 1956, on product di ff erentiation), where firms seek profits out of matchingproducts to diverse consumer profiles. In addition, utility maximisation under budget constraint is known to be an in-correct representation of attitudes towards risk, gains, losses and uncertainty (as shown in prospect theory, Kahnemanand Tversky, 1979, Tversky and Kahneman, 1974).The shortcomings of equilibrium models in accounting for multi-agent interactions and their self-reinforcing pro-cesses (complexity) or for diversity (agent heterogeneity) leads to the exclusion of very important features of realityfrom the analysis. Indeed, assuming that a simple unique equilibrium solution exists to the CBA of climate changemitigation comes at a price. It neglects both the path-dependency produced by self-reinforcing phenomena (whichare at the roots of technology lock-ins, financial bubbles and crises, technology transitions, etc.) and the diversity ofagents (which determines the rates of adoption of innovations). Such analytical omissions are reflected in the typesof policies that are advocated on that basis. Specifically, they lead to the expectation that a simple internalisation ofenvironmental costs, in the form of a pricing instrument (e.g. a carbon price or tax) will optimally and e ff ectivelyincentivise technology adoption and di ff usion. And, standard CBA fails to explain why, in the real world, such instru-ments do not play out and deliver as expected (Grubb et al., 2014). In point of fact, the expected optimal success ofsuch mechanisms is but a reiteration of the initial assumptions of the model, rather than a result of the actual analysis:e ffi cient markets, rationality, etc.Compared to equilibrium models of the economy, complex, path-dependent models may appear to be less straight-forward to interpret, but should ultimately be easier to relate to reality. Complexity is routinely handled in climatologysimulations, where it is well understood that small variations between model runs in their starting values (e.g. pres-sure, temperature, wind velocity) lead to large di ff erences in model outcomes (rainfall, cloud cover, etc) that increaseexponentially with simulated time span. This is due to a high degree of non-linear interaction between variables. Thisaspect is very well characterised and expressed as probability distributions for climate impacts (e.g. in the IPCC,2013a, summary for policy-makers) Uncertainty increasing with time of projection will arise in almost any domainwhere complex interactions between system components exist, not least in economics. A key purpose of this articleis to argue that economic analysis could benefit from harmonising its methodology with that of the climate sciences. Integrating both complexity and behavioural sciences as applied to economics would bring path-dependency andagent heterogeneity to the core of the analysis. Complexity science is the cross-disciplinary field that specificallystudies properties that emerge from interactions between system components, initially studied in physical, biologicaland computer sciences (e.g. Anderson, 1972, Sigmund, 1993). By introducing descriptions of how agents behaveand interact, theories and models turn from normative to positive / descriptive and their methodology evolves fromoptimisation to simulation, where the analyst relies on ‘what if’ approaches.It is understood that in simulations of complex systems, uncertainty plays an important role partly because thetheories do not necessarily predict a propensity to return to equilibrium. This is an aspect well understood andmanaged in the climate sciences, and no inherent reason prevents us from using the same concepts in economics.Indeed, humans behave in unpredictable ways, partly dictated by diverse contexts. The fact that humans have agencyor ‘free will’, is a frequent objection raised by social scientists, to the modelling of behaviour. This argument doesnot contradict our approach, however, for the following reason. Humans do not behave randomly (unlike physicalparticles); however, while the actions of individuals cannot be predicted, they almost always lie within known bounds,over which statistics can be developed when considering large groups of people (as for physical particles – e.g. the4ultinomial logit in social sciences is conceptually identical to the Boltzmann factor in statistical physics). The resultis a theory of collective behaviour. In complex systems (e.g. the climate, the economy), both natural and socialsystems face the same complexity challenges, and their description can be modelled in the same way.In practical terms, the complexity-based methodological approach proposed in this article follows an analyticalstructure where many scenarios are simulated based on possible policy choices, and acceptable results are retainedbased on a multidimensional range of outcomes that reach given objectives. For example, such an approach caninvolve filtering scenarios, in multidimensional human and biogeochemical space (e.g. multiple indicators such asthose likely to be selected for the recently adopted Sustainable Development Goals), to determine ranges of policyoptions that enable society to avoid exceeding planetary boundaries (Rockstr¨om et al., 2009, Ste ff en et al., 2015)while ensuring continued human development globally (OXFAM, 2012). Ambiguity or conflicting perceptions ofoptimal policy-making between diverse policy-makers or model users is avoided by leaving subjective outcome valuejudgments outside of the scientific framework. Indeed, unlike conventional multi-criteria analysis, where optimisationis carried out using subjective weighting of all factors considered given by the modeller, in the context we proposeone obtains feasible points in a multidimensional outcome space, and the policy-maker can target a possible subspacein accordance with her / his political platform. A sustainability transition inherently involves socio-technical change, which is a highly non-linear, self-reinforcingprocess with lock-ins that drive expectations, propelled by choices of and adoption by diverse agents with di ff erentperspectives and incomes (e.g. Geels, 2002, Rogers, 2010). For such a system, complexity and behavioural sciencesprovide a suitable analytical framework. Indeed, technology transitions have an inevitably significant influence onthe evolution of the economy through productivity and structural change, which occurs with the development ofnew firms and industries, and the destruction of others (e.g. Arthur, 1989, Freeman and Louc¸ ˜a, 2001, Perez, 2001,Schumpeter, 1934, 1939). At the same time, an economic transformation towards higher or lower sustainability takesplace in direct interaction with the environment and its biogeochemical cycles. Simultaneous representations of thesefour domains (technology, society, the macroeconomy and the environment) cannot reliably be optimised even on thelargest super-computers. This is so partly because diverse agents will have conflicting definitions of optimality.By contrast, in models capable of accounting for agent diversity and multi-agent interactions, interactions amongtechnology, society, the macroeconomy and the environment can indeed be simulated, much like in models used tosimulate the climate. Such models would need to be applied within a framework of uncertainty analysis, in that theywould determine the likely outcomes (within uncertainty bounds) of di ff erent policies or policy packages as applied tointeracting heterogeneous agents. Such knowledge would provide policy-makers with a much more realistic platformto make decisions, and a platform that could be more easily tailored to the specific features of a given policy context(e.g. a region, a country, a sector, etc.).A complexity paradigm has been suggested as a practical approach to policy problems (Probst and Bassi, 2014).A field of research is also emerging for modelling socio-technical regime transitions (e.g. Holtz 2011, Holtz et al.2015, Mercure 2015, Safarzynska and van den Bergh 2010, 2012, K¨ohler et al. 2009, see also Turnheim et al. 2015),in which ‘core characteristics of transitions’ are itemised as (i) ‘multi-domain interactions’, (ii) ‘path-dependency’,and (iii) ‘drivers and self-reinforcement of change’ (Holtz, 2011). System Dynamics (Sterman, 2000) and agent-basedmethods (Buchanan, 2009, Tesfatsion, 2002), however, currently have scalability challenges and thus have not yet beenintegrated into IAM scale analysis, or into other forms of global macro models. It is not necessary – in principle – tomodel every individual agent across sectors and its interactions. Equivalent but simpler statistical models, for instanceof technology adoption by interacting agents, can be readily scaled up (Mercure, 2012, 2015, Mercure et al., 2014),and non-equilibrium models of the global economy already exist (e.g. Cambridge Econometrics, 2014a, Meyer andLutz, 2007). At the next level of complexity, representative cohorts or ensembles can be simulated to obtain systemstatistics, while still simulating orders of magnitude fewer individuals than in the real system. Such approaches areused successfully in global ecosystem models such as HYBRID (Friend and White, 2000) and LPJ-GUESS (Smithet al., 2001) as well as in intergenerational investment and savings modelling (Gokhale and Kotliko ff , 2002, Miles,1999, Rangel, 2003).In this paper, we provide some concrete examples of how complexity and behavioural sciences can be used forthe assessment of sustainability policies, with emphasis on model uncertainty analysis, in order to build a powerfulapproach for next-generation public policy analysis. We identify four key areas of climate policy analysis where5ncertainty is high and where the proposed methodological approach would be particularly useful: dynamics of tech-nology adoption and di ff usion (section 2.1); macroeconomic impacts of low-carbon policies (section 2.2); interactionbetween human and environmental systems (section 2.3); and policy implementation and e ff ectiveness (2.4). We thenprovide four specific applications of the proposed methodological shift in connection with: the di ff usion of transporttechnology (section 3.1); impact of low-carbon investment on income and employment (section 3.2); processes withcascading uncertainty (3.3); and cross-sectoral impacts of biofuels policy (3.4).
2. Four key areas of climate policy where uncertainty is high ff usion Both complexity and agent heterogeneity are important to analyse the dynamics of technology adoption and di ff u-sion. The diversity of consumers and investors influences the choice of environmental technology. Consumers’ choiceof certain technologies, such as vehicles, takes place within contexts of distributed income that span several orders ofmagnitude (e.g. Mercure and Lam, 2015). The diversity of incomes, social groups and attitudes is known to determinethe rates and profiles of di ff usion of innovations (early and late adopters, etc, see Rogers, 2010). For present purposes,that means that agent heterogeneity matters in real life, and it must therefore be integrated into models. Modelling theimpacts of policy incentives to the consumer using averages over too few consumer group parameters can indeed bemisleading, since di ff erences in income and other socio-economic parameters across consumers lead to highly diverseconsumption habits.Meanwhile, interactions between agents are also crucial to consider when analysing technology adoption, sincetechnology generally features increasing returns to adoption : the adoption of technology by agents is increasinglymore probable the more agents adopt and use the technology. Transitions build upon themselves with dynamicsfollowing a pattern comparable to that of infectious diseases (see Mansfield, 1961). Thus, increasing returns maygenerate resistance to change in socio-technical regimes (lock-ins), but they may also give momentum to technologicaltransitions (Anderson et al. 1989, pp 9-10, Arthur 2014, 1989). This aspect is treated in detail in the traditions ofevolutionary economics (e.g. Freeman and Perez, 1988) and transitions theory (e.g. Geels, 2002).For obvious reasons, anticipating how diverse agents may respond to di ff erent policies has elicited great interest(see Grubb et al., 2014). The impact of incentives to consumers with di ff ering incomes and social groups has beenstudied in detail from the perspective of marketing (empirically, e.g. Bass, 1969, Fisher and Pry, 1971, McShaneet al., 2012), anthropology (Douglas and Isherwood, 1979) and behavioural economics (discrete choice modelling,e.g. Ben-Akiva and Lerman 1985, Domencich and McFadden 1975, and behavioural response, e.g. Kahnemanand Tversky 1979, Tversky and Kahneman 1974). However, the contribution of this research to the understanding ofclimate policy has so far been mostly overlooked in climate change mitigation modelling or environmental assessmentresearch. Their use thus remains a promising but emerging field (e.g. Axsen et al., 2009, Giraudet et al., 2012, Riversand Jaccard, 2006, K¨ohler et al., 2009). Grubb et al. (2014) emphasises the need for future research on behaviouralaspects of emissions reductions. Insu ffi cient e ff orts have been devoted to understanding the aggregate behaviouralresponse to sustainability policy instruments, and this is related to a possibly overstated expectation that an e ffi cientequilibrium should emerge on its own in technology markets.Empirical marketing research literature shows that technology substitution in many di ff erent contexts followssimple S-shaped di ff usion profiles (e.g. Fisher and Pry, 1971, Gr¨ubler et al., 1999, Mansfield, 1961, Marchetti andNakicenovic, 1978). More generally, the competition between several technologies for market space can be describedby coupled Lotka-Volterra systems (Bhargava, 1989, Karmeshu et al., 1985). More recently, it has been shown thatLotka-Volterra systems can be derived from simple statistics of industrial dynamics (Mercure, 2015). This can equiv-alently be expressed with the ‘replicator dynamics’ system of evolutionary theory: natural selection is carried out bythe consumer, who filters successful innovations based on their fitness to the market, while entrepreneurs strive to im-prove their products in order to increase their fitness by better matching consumer taste (e.g. Safarzynska and van denBergh, 2010). This natural selection, however, involves decision-making by consumers under bounded rationality,who are naturally highly diverse, and the diversity of consumers drives product di ff erentiation and increasing productdiversity (for example with private vehicles, see section 3.1 below).Equilibrium / optimisation models neglect agent heterogeneity and interactions, as they look at average agents inisolation facing choices (with perfect information), rather than at socio-technical systems where agents influence6 r e qu e n cy f j ( x) f i ( x)x i x j Consumer choices
Generalised cost x σ i σ j Generalised cost difference Δ x C u m u l a t i ve F r e qu e n cy agents adopting j agents adopting i F ij ( Δ x) σ ij = σ i + σ j = 1- F ji ( Δ x) Figure 1: Schematic representation of a binary logit, reproduced from (Mercure, 2015). Agent perceptions of the generalised cost (i.e. includingnon-pecuniary terms) of choice options vary around central values. In a choice between two options with distributed costs, neither is adopted byall agents; each gets a fraction of agent choices according to a comparison between perceived cost distributions, and the rate of adoption dependsstrongly on the diversity of agent perceptions. what other agents do. Agent heterogeneity could, however, be represented to some extent in models by introducingstatistical distributions over agent perspectives (see Figures 1 and 3: preferences are often simply distributed), insteadof using mean values only. In such a framework, the modeller disposes of both mean values as well as ranges. Then,when evaluating aggregate agent propensities towards technology substitution, one simply faces comparisons betweenprobability distributions, as for instance in a binary logit. Chains of binary logits enable to define agents with boundedrationality and to model di ff usion dynamics, not typically done in current models (Mercure, 2015).This is an area where our methodological shift may advance the understanding of sustainability policy by movingthe research focus from the generation of ‘desirable’ energy sector storylines (policy formulation) to the forecasting of‘likely’ policy outcomes (policy assessment) based on existing technology, knowledge of the market, and technologydi ff usion dynamics. More fundamentally, the proposed methodology disentangles normative and descriptive analysisas useful but distinct perspectives (see Mercure et al., 2014, for an extensive discussion). However, as is the casewith any model involving non-linear dynamics, uncertainty over model outcomes due to uncertainty over parametersincreases exponentially with modelling time span, an issue that needs to be addressed (see section 2.3 below). Complexity and agent heterogeneity are also important to understand the interactions between technology dif-fusion, lenders’ expectations and macroeconomic fluctuations, which drive economic development / growth and area ff ected by climate change mitigation policies. Theories based on utility-maximising non-interacting agents assumefull employment of economic resources. Under full employment, policies that lead to a re-allocation of resourcesfrom an optimal equilibrium starting-point are unavoidably detrimental to economic performance. However, full em-ployment is never observed in reality. In real life, expectations as to the performance and return of innovations drivemost investment, and ultimately, economic development and growth. Importantly, expectations as a process arise frommulti-agent interactions (e.g. see the artificial stock market by Arthur et al., 1997).Policies for the di ff usion of low-carbon innovations present significant opportunities for both the private sectorand job creation (e.g. Blyth et al., 2014, CBI, 2014). But the feedback between the di ff usion of new sustainabilitytechnologies and economic development / growth is poorly understood and reported. There is an apparent contradictionbetween observations of highly lucrative activity arising in successful low-carbon ventures (e.g. Tesla electric andToyota hybrid cars, wind turbines, solar photovoltaic), and the perception that the use of lowest-cost fossil energywith conventional technology is indispensable for development / growth. This type of dichotomy pervades climatepolicy-making. 7n conventional equilibrium economics, full employment (which results from maximising the utility of the repre-sentative agent) entails that all resources are currently allocated in the economy in the most productive way they canbe. This is contradicted by empirical observation (e.g. as reviewed by Grubb et al., 2014), which shows that di ff erentcountries lie at varying distances from theoretically defined e ffi ciency or productivity frontiers. E ff ectively, the use ofeconomic resources depends upon the economic development trajectory followed by an economy (investment, fixedand human capital and labour, ibid ). Moreover, in equilibrium-based theories, savings are understood as a share offixed national income (GDP), equal to investment, resulting from a choice by agents between spending on consump-tion in the present or in the future. Thus if a fixed share of income is allocated to investment, naturally, all firmscompete for this fixed allowance (crowding-out), equilibrating supply and demand for finance. Since in equilibriuminvestment and labour is fixed, equilibrium-based theories cannot reproduce financial fluctuations or crises, such asthose observed since 2007, or involuntary unemployment, as currently observed in many parts of Europe (Eurostat,2015).In a broader complexity / non-equilibrium economics perspective, income is not fixed, and thus investment doesnot need to be a constant share of the ‘welfare of the representative agent’. Instead, causality is reversed: incomeheavily depends upon investment, and investment decisions depend upon decisions by financial institutions. This link,however, opens the door to fluctuations: interacting investors in technology ventures can influence each other intofrenzies or panics. Indeed, the finance of technology and innovation leads to productivity change (a phenomenondemonstrated already in the 1950s by Solow, 1957) but also to speculation and bubbles (Keynes, 1936), which, inturn, create economic cycles (see Freeman and Louc¸ ˜a, 2001, Perez, 2001, Schumpeter, 1934, 1939). Investmentfluctuations heavily influence the level of economic development. This dynamic was at the roots of the great recession(Keen, 2011), after which quantitative easing was used to o ff set the reluctance of financial institutions to providecredit. This dynamic cannot be understood without taking into account the interactions between technology di ff usion,lender expectations and macroeconomic fluctuations.More realistic economic modelling that allows for variable amounts of finance and investment is possible if oneabandons the restrictive assumptions of equilibrium approaches (imposing a theory of non-interacting agents thatare all simultaneously employed or not employed) with their assumed full employment of economic resources. In atheoretical approach that allows for crowd e ff ects to emerge, the amount of investment in the economy is determinedby investment decisions by lenders acting upon expectations of return. This is not a new theoretical approach. Inpoint of fact, it is at the roots of both Schumpeterian (Schumpeter, 1934, 1939) and Keynesian (Keynes, 1936) formsof demand-led economics. A post-Schumpeterian or post-Keynesian perspective also allows for fluctuating amountsof money in the economy, and indeed, this is what is observed (called endogenous money , as explained by the Bank ofEngland, see McLeay et al., 2014).As understood in the tradition of Schumpeterian economic history, the emergence of new technologies is char-acterised by important increasing returns to investment and adoption, and it involves strongly path-dependent cross-sectoral spill-overs that often lead to economy-wide activity accelerators (Freeman and Louc¸ ˜a, 2001). For example,in the industrial revolution, new textile machinery required better iron and steel. Investment in iron and steel led tothe emergence of an industry as well as cost reductions that reshaped the whole ‘ design space ’ for other products tobe made out of those materials more cheaply ( ibid ). This, in turn, enabled many related and clustered innovations toemerge and economically reach the marketplace (a statistical analysis of the process of clustering of innovations isgiven by Arthur and Polak, 2006). In the Schumpeterian economic tradition, the clustering of investment stemmingfrom the clustering of innovation leads to both economic prosperity and depression periods in alternation (Freemanand Louc¸ ˜a, 2001, Perez, 2001, Schumpeter, 1939). Risks of financial crises arise when finance is raised using financialassets as collateral (Keen, 1995, 2011), a phenomenon known to have taken place periodically over history (Perez,2001), including in recent years (Keen, 2011).Formal modelling of path-dependent, non-equilibrium macroeconomics is possible on the basis of two assump-tions. Firstly, under the principle of cumulative causation of knowledge accumulation , productivity change is de-scribed by sets of aggregate sectoral learning curves where productivity growth is derived from cumulative investmentand is roughly consistent with the trends observed in economic history, e.g. in Kaldor (1957), Schumpeter (1934).Secondly, income must be allowed to depend upon investment, instead of the reverse. This understanding is path-dependent and implicitly integrates multi-agent interactions by allowing increasing returns: learning (i.e. knowledgeaccumulation), investment fluctuations (through expectations), economic cycles and technology di ff usion (through so-cial interactions). One example of such a model is E3ME / G (see , Cambridge Econometrics, 2014b),8
UIDEPOST
Model input Emulator 1 Emulator 2 Emulator 3 e.g. emissions e.g. carbon cycle e.g. climate e.g. land productivity a. b. c. d.
Figure 2: Schematic representation of cascading uncertainty across statistical model emulators. which derives a closed set of functional relationships using regressions on economic data, and thus enables the eval-uation of employment, instead of remaining constrained by the full employment assumption. Other similar modelsexist, including GINFORS (Meyer et al., 2013). However, so far, no large-scale model has included detailed dynamicsof the financial sector (Pollitt and Mercure, 2015).With regard to climate change mitigation, it is well established that an amount of investment in the energy sectorlarger than in a business-as-usual scenario will be necessary (e.g. IEA, 2012). This scale of investment in new technol-ogy sectors is likely to result in substantial economic reallocations due to cross-sectoral spill-overs (e.g. new materials,new design and engineering methods, etc). Importantly, in models without the constraint of full-employment of re-sources, capital and labour, the impact can be either beneficial or detrimental, depending on many factors, such asthe trade balance, energy imports, international competition and relative prices. When the implications are beneficial,the transition can be termed as green growth , where new investment and employment are generated from technologypolicy, although possibly with important relative price changes. In section 3.2, we discuss an example where climatechange mitigation policy leads to increased employment and GDP as a result of enhanced investment in the electricitysector. This said, technology policy does not necessarily lead to beneficial impacts because such policies can takemany forms. In an equilibrium-based theory, however, the possibility of beneficial outcomes is entirely ruled out, andby assumption rather than as a result of calculations. Indeed, equilibrium models always predict detrimental impactsfor climate mitigation e ff orts (e.g. as in all the models chosen in the 5th assessment report of the IPCC, see IPCC,2013b, Ch. 6, p.450), such that the debate is framed in terms of burden , and economic opportunities of climate poli-cies are not extensively examined (Grubb et al. 2014 see, however, Ekins et al. 2011, Lee et al. 2015). This is also thecase of partial equilibrium models, where, such as in the case of the World Energy Model utilized by the InternationalEnergy Agency for its World Energy Outlook report, the investment cost of interventions is estimated, but it is notcompared to policy-induced avoided costs nor the impact of these interventions on the macroeconomy. The result is apartial analysis that emphasizes only the costs, and not the benefits of intervention (UNEP, 2011).Thus, and importantly, we see that the incorporation of multi-agent interactions (by accounting for feedback loops,delays and nonlinearity) leads to key consequences for theory and its application: in a theory where investment isdetermined by expectations, information flow between variables runs from variable or fluctuating savings / investmentto income, back to investment, and allows for beneficial impacts of technology policy while in equilibrium-basedtheories, information flows from fixed income to investment, and beneficial impacts are ruled out. Given the scaleof the socio-economic transformation required to address climate change, a realistic understanding of whether asustainability transition involving mitigation policies favours or hinders economic development / growth is key forsound policy-making. Models based on complexity theory can better describe the complex interactions between the socio-economic andthe natural (environment) systems. Such interactions are at the heart of climate policy, and they can be e ff ectivelysimulated at a low cost through the use of statistical emulators.9eedbacks between the economy and the natural environment take place through direct human intervention (e.g.land-use change) or indirectly through the impact of economic activity (e.g. the generation of air pollution or theemission of greenhouse gases causing climate change). The natural environment influences, in turn, human actionleading to a highly complex system of feedbacks. The allocation of land for agriculture, including crops for biofuels,interacts directly with the climatic system through phenomena such as deforestation, land-use change emissions anddesertification, but also indirectly through the economy. The extensive use of water further constrains the availabilityof natural resources for human use, in what is increasingly referred to as the food-energy-water nexus . At the sametime, natural resource use stems directly from economic processes, involving the demand for agricultural, energy andforestry commodities that are traded in international markets.Exploring complex interactions between the environment and the socio-economic system requires detailed rep-resentations of the role and behaviour of the natural environment, which is no minor endeavour. Indeed, climatemodelling is carried out with the most powerful supercomputers currently available. The response of the environmentto anthropogenic influence is increasingly well understood (IPCC, 2013a) and some of this knowledge has been re-produced reliably enough through the use of emulators (statistical representations). Emulators can be used directly,without requiring detailed parallel simulations with supercomputers alongside an economic model. Such statisticalrepresentations, which interpolate climate responses based on large climate model output data, have now been usedfor some time (e.g. Meinshausen et al., 2009), facilitating the access by social and economic modellers to quantitativeresults from climate science, usually carried out in di ff erent institutions.In this context, a frequently used tool is the linear approach of ‘pattern scaling’ (Tebaldi and Arblaster, 2014).This approach assumes that the spatial pattern of change in the climate output of interest is invariant with respectto time and forcing. This approximation is often inadequate, however, for instance in the case of land-use change,which not only has significant impacts on the global climate through greenhouse gas emissions but also has a morelocalised (i.e. ‘pattern changing’) impact on climate through changes in surface albedo, moisture transfer and riverruno ff (Myhre et al., 2013). A more general approach that allows for changes in the pattern of climate impacts hasbeen developed recently (Holden and Edwards, 2010) and applied to ‘emulate’ complex models of a range of climaticsystems. The technique has been used to produce emulators for the climate model PLASIM-ENTS (Holden et al.,2014), the carbon cycle model GENIE (Holden et al., 2013), and the land surface model LPJmL (Oyebamiji et al.,2015). These emulators have already been applied in a range of integrated assessments (including but not limited toIEA, 2013, Labriet et al., 2013, Mercure et al., 2014).Such a methodology provides a useful platform for the integration of large amounts of environmental data intosocio-economic and policy analyses. By concatenating such emulators in ‘chains’ (see schema in Figure 2), one canindeed obtain a representation of uncertainty cascading across sectors. This technique avoids possible biases thatmay emerge from the use of median trajectories when linking di ff erent models. And it enables the exploration ofrepresentations of all likely trajectories simultaneously, at low computational costs.The cascading of uncertainty that increases with simulation timespan stems from the general property of complexnon-linear models, which produce scenarios that diverge from each other exponentially for arbitrarily small changesin starting parameters (i.e. the sometimes termed ‘butterfly e ff ect’ in chaos theory), typical of climate models. Thistype of uncertainty is also a property of an economic model with positive feedbacks, as well as of di ff usion-basedtechnology models. More generally, it would be inconsistent to treat economic projections as deterministic (e.g. aGDP trajectory in equilibrium), while treating climate projections as probabilistic (e.g. 95% probability range for aglobal warming trajectory). It is, however, entirely possible to assign ‘uncertainty bounds’ to projections of a socio-economic model through the use of statistical analysis. Thus, the uncertainty faced in climate policy-making can bee ff ectively tackled through such statistical shortcuts. Depending on the sensitivity of each sub-system to perturbations(i.e. the rate of divergence of scenarios, and conversely, the possible presence of ‘attracting states’) the cloud ofuncertainty may increase moderately or dramatically as it propagates through the chain. This in turn provides a clearquantification of our ability to reliably model the ensemble of systems. ff ectiveness, behaviour and implementation Tools from marketing research, anthropology and behavioural economics can be very useful to improve our under-standing of consumer and investor behaviour. They provide a powerful addition to economy-environment modelling,namely the ability to predict quantitatively the most likely aggregate response to policy instruments. Valuable in-sights can be gained in this way. In climate change policy, which requires action sooner rather than later, there is10ignificant value in the ability to predict the e ff ectiveness of emissions reduction measures. However, assessing suchfeasibility requires knowledge of their legal and political implications and constraints (e.g. political feasibility, legalconsistency), which may be just as important. Apparently minor di ff erences in the applicable legal framework (e.g.the modalities for the acquisition of land, the modalities for the initial distribution of emission reductions or for theirbanking, the local content requirements added to a feed-in-tari ff scheme to make them politically palatable) may, infact, make certain developments less e ff ective, or more vulnerable to challenge, or even block them entirely.Much policy analysis has been carried out in connection with specific policies proposed to address externalitiessuch as pollution or environmental degradation, including taxes and other financial incentives imposed on households,firms or consumers. The justification for taxes is often based on considerations of ethics and social justice (IPCC,2013b, Ch. 3); it less frequently results from the analysis of their likely e ff ectiveness (IPCC, 2013b, Ch. 15). Indeed,policies are often chosen and adopted without much prior knowledge of their likely e ff ectiveness, which ultimatelylies with investor and consumer decisions. Their impact is mostly assessed ex-post ; however here, relying on ex-postpolicy evaluation does not fit in the timescale of action.Policy analysis for climate change mitigation is complex, and realistic proposals must take into account the polit-ical and institutional (legal – both domestic and international) context. For example, some environmental di ff erentia-tion techniques (e.g. certain subsidies or feed-in tari ff s), may be inconsistent with international economic law and canattract significant legal di ffi culties (Vinuales, 2012). Furthermore, the introduction of carbon pricing mechanisms (e.g.a cap-and-trade system where the allowances are freely allocated – at least initially – on the basis of prior emissions)can be politically very di ff erent from that of energy / fuel standards or emissions taxation, and it would therefore havea di ff erent likelihood of success. Thus, studying the e ff ectiveness and feasibility of policy measures requires integrat-ing expertise on the mechanisms of policy-making and law with expertise on the modelling of decision-making byinvestors and consumers. However, such a broad integration of expertise is extremely rare in contemporary state-of-the-art climate change mitigation research.We propose here, as part of the working structure of simulation-based sustainability policy assessments, a two-wayfeedback with knowledge of domestic / comparative politics and environmental, investment and trade law. This woulde ff ectively guide the ‘what if’ approach to scenario-creation for testing potential policies, as opposed to proposingpolicies that are already ‘optimal’. E ff ectively, in a model that is not based on optimisation, policies are assessedon the basis of their ability to e ff ectively achieve certain objectives through the simultaneous use of several policyinstruments that interact with one another. Such an approach avoids the more common siloed assessments, and itaccounts for policy impacts across sectors. This approach is in fact recommended by the European Commission in itsImpact Assessment guidelines (EC, 2009, 2015).
3. Practical relevance of the paradigm shift: specific applications to four key climate policy issues ff usion of new transport technology The importance of agent heterogeneity for technology adoption can be empirically established, as shown in fig-ure 3, using private passenger vehicle purchases as an example derived from recent work (Mercure and Lam, 2015).As shown in figure 3, in the UK, the distribution of car purchase prices follows closely the income distribution (panela). Cars of di ff erent prices have di ff erent manufacturer-rated emissions, which are thus similarly distributed (panelb). Vehicles with alternative engine technologies have yet a di ff erent distribution, and this stems from their gradualprocess of di ff usion, which takes place unevenly with respect to the distribution of conventional vehicles (panel c).Finally, rated emissions are correlated with vehicle prices through a log-linear relationship.Car purchase choices largely depend upon each consumer’s respective social group, through social interactions,as has been shown in empirical work (for example McShane et al., 2012). This generally explains the relationshipbetween consumption behaviour and income distribution shown above. Consumers do not attempt to minimise theirtransportation costs; instead they apparently purchase what is most common in their visible surroundings, and thediversity of social groups is what forms the lognormal distribution shown above. Alternatively, we can say thatconsumers maximise utility within a subset of the market (bounded rationality) defined by their social interactions.It thus appears necessary to look at each band of income, and each type of consumer, separately. If we now explorepossible substitutions between models that could result from emissions reduction policies in the car market, we cancalculate elasticities of substitution from these distributions by determining statistically which vehicle models are most11
20 40 60 80 100020406080100120140 Car Price (kUSD) S a l e s ( t hou s and s ) UK All vehiclesHybrid x 10Electric x 100 b) % popu l a t i on UK income distributionUK car purchasesdistribution a)
10 20 50 100 200050100150200250300350400 Vehicle Price ( k2012USD ), log scale E m i ss i on s ( g C O / k m ) R = 0.71R = 0.50R = 0.63Petrol EconPetrol MidPetrol LuxDiesel EconDiesel MidDiesel LuxHybrid EconHybrid MidHybrid Lux 100100010 000Sales numbers d) /km) S a l e s ( t hou s and s ) UK S a l e s ( t hou s and s ) All vehiclesHybrid x 10 c) Figure 3: Data relating the impacts of consumer diversity and market structure on the e ff ectiveness of emissions taxation in the private personalmobility sector (reproduced from Mercure and Lam, 2015). a) Comparison of the UK income distribution to the price distribution of recent vehiclepurchases. b) Comparison of UK market coverage between conventional and unconventional vehicle engine technology. c) Associated distributionof vehicle carbon intensities. d) Structure of the UK vehicle market for prices and emissions from which likely e ff ectiveness of taxes can bededuced. These properties were found to vary across economies (Mercure and Lam, 2015). likely to be chosen within the same new price band (after tax). Therefore, the rate of adoption of consumer technology,including low carbon systems, stems precisely from this diversity, which varies across the world (as shown in Mercureand Lam, 2015). This enables us to determine the e ff ectiveness of certain low-carbon policies (taxes, subsidies) atincentivising technology substitution, using market data. In point of fact, this is exactly what marketing research doesto forecast sales before placing new products in the market.Such information can be fed into technology di ff usion models that aim to reproduce typical S-shaped profiles (e.g.as in the Future Technology Transformations (FTT) model, Mercure, 2012) to provide expected rates of adoption (seeFigure 5 further below). Such a representation reproduces technology lock-ins. This picture is incomplete, however,as it does not represent attitudes and culture. Yet, just as with firms attempting to place new products in the market,the more detailed information thus gained helps to better characterise the rates of technology adoption that couldresult from proposed policies throughout the di ff usion cycle. By contrast, optimisation-based models would typicallycharacterise variations in consumer behaviour, at best, by parameterising di ff erent discount rates for technologiesattributed to particular market segments. The importance of accounting for the interactions between lenders’ expectations, technology di ff usion and macroe-conomic variables can be illustrated by reference to the link between low-carbon investments, income and employ-ment. Figure 4 is based on a non-equilibrium macroeconomic model, which is not subject to the full employment12 D i ff e r en c e ( bn$ ) a . Investment (2000bn$) 0510152025 b . Gov. Spending on subsidies 050010001500 c . Gov. income from ETS0204060 D i ff e r en c e ( % ) d . Electricity prices − e . Employment 051015 f . Real disposable income2010 2020 2030 2040 2050 − D i ff e r en c e ( % ) g . Consumption 2010 2020 2030 2040 2050 − h . Consumer price index 2010 2020 2030 2040 2050 − − i . Real GDPNorth AmericaEurope ChinaIndia BrazilRest of the World Figure 4: Example of possible economic impacts of investment in low-carbon electricity generators, when studied using a non-equilibrium post-Keynesian model, E3MG-FTT (see Mercure et al., 2015). Vertical axes refer to changes from the baseline, in billion current USD for the top threepanels, and in percentage points for the six lower panels. constraint (E3ME / E3MG Cambridge Econometrics, 2014b). It shows a possible causal chain in the process of tech-nology finance based on expectations impacting the economy in the electricity sector. This calculation is furtherdescribed in Mercure et al. (2015).In this model, no limit is imposed on finance for technology developments, which in contrast to an equilibriummodel, has no direct relationship to interest rates (assumed in equilibrium to clear the money market). In otherwords, financial resources are given to entrepreneurs by banks for low-carbon investments through credit creation.The model’s critical assumption is only the solvency of entrepreneurial activity, i.e. economic viability of low-carbonprojects. Therefore it is not claimed that the economic system has no limit at all on money flows: it is assumed that alltechnology ventures that are financed in any scenario are profitable (for instance by assuming sustained credible policyand / or prices that make these ventures feasible), and that government and / or private debt is not indefinitely increased.This ensures that situations that could lead to financial instability through unsustainable debt growth are not created.This example is selected to show that economic outcomes of mitigation policy in models are entirely dependent onmodel architecture, and that impacts can be (but are not necessarily) beneficial if one allows for non-equilibriume ff ects.In this example, the electricity sector is decarbonised by 90% (scenario previously published in Mercure et al.,13014). Higher costs of low-carbon electricity generation technologies (e.g. wind turbines, solar panels, carbon captureand storage) are passed on by utilities into their bills to customers, i.e. into higher prices of electricity, proportionallyto the evolving technology composition. Entrepreneurial activity requires loans to finance new low-carbon capital,which requires more investment than existing fossil fuel generators. Banks create money to finance these ventures .Loans are paid back over capital lifetime by firms using a higher price of electricity (and, for example, possible feed-intari ff s), and the additional money is gradually destroyed as the loans are paid back. In the model, as shown in figure 4,the higher level of money flows from investment (panel a) creates jobs (panel e), increases disposable income (panelf) and consumption (panel g), with possible e ff ects on inflation (panel h). Meanwhile, higher prices of electricityincrease operational costs of most sectors, and thus decrease employment, household income and consumption (samepanels), i.e. an o ff setting force. These two e ff ects were observed to roughly cancel each other in the model andscenario (see Mercure et al., 2015). A positive ‘green growth e ff ect’, in this particular case, is generated in parts dueto redistribution policy, in parts due to increased employment. Revenue raised by fuel taxes (panel c) aimed at incen-tivising technological change (carbon pricing), minus spending on technology subsidies (panel b), is indeed recycledto lower income tax. This moves the system’s balance towards a higher income level than in a baseline scenario(higher GDP, panel i). This e ff ect subsides in later years (US, Europe) when investment and redistribution declines,while the price of electricity remains high, and the e ff ect may even reverse when the technology transformation iscompleted and society faces servicing debt only. Debt servicing takes place through consumers paying a higher pricefor electricity.Note, therefore, that while economic growth is enhanced in this scenario, private debt is also increased beyondthe end of the simulation. In non-equilibrium theory, borrowed investment flows indirectly contribute to increasedaggregate demand in the short run, when loans are issued, and to decreased demand in the long run, when loansare gradually paid back. Increasing the level of borrowing generates debt-based growth, which if done indefinitely,generates significant prosperity but eventually leads to collapse through a financial crisis. It is unlikely that climatechange mitigation would lead to indefinite borrowing. But it will likely require significant amounts of finance.According to Keen (e.g. Keen, 2011), extended debt-based growth and excessive debt levels have been the under-lying cause of the recent banking crisis, and possibly many other economic cycles historically (Perez, 2001). Thispoints to a complex entanglement between strategies for economic recovery after the economic crisis, and strategiesfor climate change mitigation policy. The financial crisis involved banks refusing to lend despite quantitative easing,while mitigation, a potential economic stimulant, requires increased amounts of finance in the energy sector.We conclude that further research is required to understand the levels of debt and risk, private and public, thateconomies can realistically undertake to reduce global emissions, as well as to clarify the collective expectations ontheir economic returns. In a complexity / non-equilibrium perspective, the impact of collective expectations appearsto be the key issue to explain, rather than simple welfare e ff ects of di ff erent allocations of fixed amounts of capitaltypically analysed with current optimisation models. In the context of scenarios of rapid decarbonisation, the analysisof debt and access to finance becomes particularly crucial. Large models of the natural world have many imperfectly known parameters as well as – through multiple com-binations of those parameters – a considerable number of possible output values arising from variations in thoseparameters (even 10 settings of only 10 parameters would generate 10 possibilities). Statistical modelling tech-niques are required to interpret such a large space of uncertain outcomes. When several models are ‘soft-linked’ ina causal chain, the uncertainty of models upstream inevitably generates higher uncertainty downstream as we giveever wider ranges of parameters to models. Figure 5 gives an example of this with 3 soft-linked models, starting withindividual emissions scenarios from E3MG-FTT (note that these scenarios are the same as in Figure 4, their emissionstrajectories coupled to climate emulators; they are also the same as scenarios a. and j. in Mercure et al., 2014).In (a.), we have a baseline scenario of the composition of the global electricity sector, calculated under 21 regionsindependently. Using a composite scenario of emissions reduction policies that include carbon pricing, technologysupport policies (subsidies and feed-in tari ff s) and regulations (available in Mercure et al., 2014), the electricity sectoris transformed towards low-carbon technologies (b.). Baseline emissions are projected to increase by 318% based ontheir 1990 level (c.), while in the decarbonisation scenario, they are reduced by 90% (d.).These emissions scenarios are fed to the carbon cycle emulator GENIEem, which generates GHG concentrationoutputs with 95% probability ranges (e.). These are then fed to the emulator of the climate model PLASIM-ENTS,14
980 2000 2020 204001020304050 E l ec t r i c i t y G e n e r a t i on ( G W h ) Business As Usual
Historical data Projection1990 a. Decarbonisation b. C O c on ce n t r a t i on ( pp m C O ) E m i ss i on s ( G t ) c. d. G l ob a l W a r m i ng ( o C ) G l ob a l e m i ss i on s ( G t C O ) NuclearOil CoalCoal+CCS GasGas+CCS BiomassBiomass+CCS HydroWind SolarGeothermal Ocean
BAUMitigation e.f.
Figure 5: Example calculation of the environmental impacts of electricity policy instruments using E3MG / E-FTT (left 4 panels) with OU emulators(right panels), from policy to global warming, with cascading uncertainty bounds from combined carbon cycle and climate system emulators. producing a set of scenarios for global warming (f.), as well as other locally resolved changes of climate (on a 0.5ogrid, not shown). The soft-linking thus produces double uncertainty ranges from the concatenation of uncertainty. Wefind that decarbonising the electricity sector by 90% is not su ffi cient to avoid exceeding 50% chance of exceeding theinternational target of 2 ◦ C (Mercure et al., 2014). All sectors of the energy end-use system must be involved, notablytransport. Cascading uncertainty ranges in this context is as important as cascading median trajectories.
In an optimisation model of land allocation, best allocations of farming activities are determined across the surfaceof the land studied, which extends to the whole world for global models. This implies that farmers know instanta-neously what is too little and too much, and never generate excess product. Without excess product or shortages,price fluctuations cannot occur. This is, of course, not what is observed (Piesse and Thirtle, 2009). However, build-ing a model that can reproduce observed price fluctuations is a significant challenge, even though such fluctuationsseem crucial to understand environmental degradation and biodiversity loss. Indeed, some deforestation occurs not di-rectly because of an increasing consumption of agricultural commodities globally, but as a result of commodity pricesfluctuating globally, due to self-reinforcing expectations of returns (e.g. Arima et al., 2011, Morton et al., 2006).In a non-equilibrium model that includes representations of decision-making by heterogeneous agents in agricul-ture with expectations of return on their crops, in tandem with a global model of the economy and bilateral trade ofagricultural commodities, very di ff erent dynamics can emerge. The use of heterogeneous agent functional types hasbeen suggested to improve model realism (Arneth et al., 2014, Rounsevell et al., 2014). While price substitutionsand changes in trade patterns can occur with the consumption of agricultural commodities, massive land clearanceand conversions can also take place to accommodate expectations around changing prices. Such models generatehighly complex dynamics, but are useful to determine what types of transformations could occur in the future without15ppropriate land management policy. Complex dynamics however will only arise in models that incorporate hetero-geneity, increasing returns and expectations. When land-use decisions are based on expectations and influenced byneighbourhood e ff ects, then similar contagion dynamics as in technology di ff usion may take place, i.e. the di ff usionof agricultural practices.When brought into the picture, biofuels policy adds further uncertainty and complexity. On the one hand, thewillingness to pay for ethanol in some parts of the world could outbid the ability to pay for food commodities inother parts of the world, while, on the other hand, shortages of food may also be felt through commodity prices afterland-use decisions are already made and applied. E ff ectively, some biofuels policies have the potential to open thedoor for substantial fluctuations in food prices, but we do not know which ones, and at what scale. This issue requiresmodels of a type that is currently not available, but needs to be addressed rapidly. Lower complexity or reduced-form models are often argued to be advantageous to use due to their higher apparenttransparency, and thus more appropriate for policy-making. An example is the use of Nordhaus’ DICE model, whichwas originally designed for illustrative purposes using sketchy data, but has been used for policy-making purposes insome countries, including the USA and the UK.Lower complexity does not mean better science , and simplifications can, in some cases (e.g. as with Nordhaus,2010) lead to potentially plainly incorrect conclusions (its single sector / good representation of the economy excludesmany important e ff ects such as spill-overs and multipliers). Moreover, and more fundamentally, the use of determinis-tic model projections (e.g. policy-making based on an optimal carbon price that results from intersecting deterministiccurves for the social cost of carbon and the marginal abatement cost) is not a scientifically correct methodology. Itis clearly inconsistent with the way complexity-based climate modelling is conducted. No one would realisticallyclaim the ability to predict the exact mean global warming in 2050, and there is no reason why such claims would bejustified in economics.We thus argue that while higher complexity models may be more di ffi cult to use in policy circles, such di ffi cultyis clearly o ff set by their greater realism and rigour. The key consideration is not whether simpler or more complexmodels should be used, but whether the science-policy interface is capable of relaying the results of more powerfuland realistic models to policy circles.
4. Conclusion: a world of new possibilities for sustainability policy-making
Equilibrium and optimisation-based models are appropriate to use for normative exploration and identification ofdesirable future configurations of the technology-economy-environment system. Given the fact that they are com-paratively highly detailed and tested, they are currently treated as the standard approach. This is possible becausenormative analysis relies entirely on assumptions and does not require empirical knowledge of actual human be-haviour. However, such models support only certain steps of the policy cycle, as they provide ambiguous informationregarding how to achieve – through policy interventions – the ideal configurations they portray or how likely they maybe. This gap is a direct consequence of a lack of causal relationships with human behaviour . Producing scenarios thataccurately forecast the future course of events as a result of policy choices, with any likelihood (however low), requiresfine-grained representations of human behaviour, its diversity and multi-agent interactions. These representations areas necessarily imperfect as they are necessary to come close to real life.It is often argued that forecasting is not possible, and that serious sustainability science can only express itselfin the form of exploratory scenarios (e.g. van Vuuren et al., 2011). We submit in contrast that forecasting is bothnecessary and inevitable, and can be improved within the bounds of finite predictability with increased attentionto known nonlinearities and interaction e ff ects (Tetlock and Gardner, 2015). The climate sciences provide a majorexample of an area where outcomes are expressed in terms of likelihood levels based on model statistics. Forecasts arenot always accurate, but they are nevertheless useful due to their careful quantification of probabilities, parameters andtheir uncertainty. There is no inherent reason why this is not possible in the social sciences. We note that the climatecommunity now prefers the term ’projection’ to indicate that model simulations are predictions that are conditionalon their inputs. This distinction of terminology, while relevant to the remaining exogenous assumptions in our case,does not a ff ect our fundamental conclusion. 16he main challenge lies in the quantification of likelihood, which is not possible with current mainstream socio-economic models. We do not claim, however, to be able to predict the onset of wars, election results, natural disasters,strategic political choices or other unique events. Thus, an important caveat must be made, in that our methodologicalapproach excludes the occurrence of some events (just like in weather forecasts, which do not take into accountpossibilities of volcanic eruptions, even in active areas).Yet, the modelling paradigm shift proposed here opens a large spectrum of possibilities for sustainability policy-making. It allows for the quantification – within uncertainty bounds – of the expected e ff ectiveness of specific policiesaimed at inducing particular agent groups to take particular decisions (e.g. consumer purchases, investment choices,land-use decisions). The ability to conduct such forecasts entails significant advantages, including the reduction of theuncertainty involved in creating policy portfolios, and the possibility to explore their impacts across sectors through thecoupling of several sectoral models. For example, one could study the impact of technology support policy for electricvehicles on electricity prices, which would depend on the pace of their adoption, or the impact on food prices of largescale land regulations to protect biodiversity or, still, the impact on deforestation of biofuels policy in transport. Thesequestions present major analytical challenges, but their understanding cannot wait any longer. We believe that the newgeneration of non-equilibrium models proposed in this article is capable of rising to this challenge. Acknowledgements
The authors wish to thank the full GUIDEPOST team for highly fruitful discussions over the length of fundingproposal developments. The authors also warmly thank assiduous referees for their comments that led us to greatlyimprove our manuscript, and A. Jarvis, whose comments made us revisit concepts in detail. J.-F. M. thanks D.Crawford-Brown, T. Barker and 4CMR sta ff and students for support. J.-F. M. and H. P. thank Cambridge Econo-metrics Ltd sta ff for support. J.-F. M. acknowledge the UK Engineering and Physical Sciences Research Council(EPSRC), fellowship no EP / K007254 / / N002504 / ReferencesReferences
Anderson, P. W., 1972. More is di ff erent. Science 177 (4047), 393–396.Anderson, P. W., Pines, D., Arrow, K. J., 1989. The economy as an evolving complex system. Santa Fe Institute studies in the sciences of complexity.Westview press.Antho ff , D., Tol, R., 2014. Climate framework for uncertainty, negotiation and distribution.URL Arima, E., Richards, P., Walker, R., Caldas, M., 2011. Statistical confirmation of indirect land use change in the brazilian amazon. EnvironmentalResearch Letters 6 (2).Arneth, A., Brown, C., Rounsevell, M., 2014. Global models of human decision-making for land-based mitigation and adaptation assessment.Nature Climate Change 4 (7), 550–557.Arthur, B., 2014. Complexity and the economy. Oxford.Arthur, W. B., 1989. Competing technologies, increasing returns, and lock-in by historical events. The economic journal 99 (394), 116–131.Arthur, W. B., Holland, J. H., LeBaron, B. D., Palmer, R., Tayler, P., 1997. Asset pricing under endogenous expectations in an artificial stockmarket. In: The economy as an evolving complex system. - 2.Arthur, W. B., Polak, W., 2006. The evolution of technology within a simple computer model. Complexity 11 (5), 23–31.Axsen, J., Mountain, D. C., Jaccard, M., 2009. Combining stated and revealed choice research to simulate the neighbor e ff ect: The case ofhybrid-electric vehicles. Resource and Energy Economics 31 (3), 221 – 238.Bass, F. M., 1969. New Product Growth for Model Consumer Durables. Management Science Series A-theory 15 (5), 215–227.Ben-Akiva, M. E., Lerman, S. R., 1985. Discrete choice analysis: theory and application to travel demand. MIT press.Bhargava, S. C., 1989. Generalized Lotka-Volterra Equations and the Mechanism of Technological Substitution. Technological Forecasting andSocial Change 35 (4), 319–326.Blyth, W., Gross, R., Speirs, J., Sorrell, S., Nicholls, J., Dorgan, A., Hughes, H., 2014. Low carbon jobs: The evidence for net job creation frompolicy support for energy e ffi ciency and renewable energy. Ukerc reports, UKERC.URL Buchanan, M., 2009. Meltdown modelling: Could agent-based computer models prevent another financial crisis? Nature 460, 680–682.Cambridge Econometrics, 2014a. E3ME Manual.URL ambridge Econometrics, 2014b. Energy-economy-environment model at the european level.URL CBI, 2014. The colour of growth: Maximising the potential of green business. CBI reports, CBI UK.URL
Domencich, T. A., McFadden, D., 1975. Urban travel demand - A behavioural analysis. North-Holland Publishing.Douglas, M., Isherwood, B., 1979. The world of goods: Towards an anthropology of consumption. Routledge.E3MLab, 2013. The gem-e3 model.URL
E3MLab, 2015. The primes model.URL
EC, 2009. Impact assessment guidelines. Tech. rep., European Commission.URL http://ec.europa.eu/smart-regulation/impact/commission_guidelines/docs/iag_2009_en.pdf
EC, 2015. Better regulation guidelines. Tech. rep., European Commission.URL http://ec.europa.eu/smart-regulation/guidelines/docs/swd_br_guidelines_en.pdf
Ekins, P., Pollitt, H., Barton, J., Blobel, D., 2011. The implications for households of environmental tax reform (etr) in europe. Ecological Eco-nomics 70 (12), 2472 – 2485.Eurostat, 2015. Eurostat Statistics, Employment and unemployment (LFS). Eurostat.Fisher, J. C., Pry, R. H., 1971. A simple substitution model of technological change. Technological Forecasting and Social Change 3 (1), 75–88.Freeman, C., Perez, C., 1988. Structural crises of adjustment: business cycles and investment behaviour. In: Dosi, G., Freeman, C., Nelson, R.,Silverberg, G., Soete, L. (Eds.), Technical change and economic theory. Pinter Publishers, pp. 38–66.Freeman, C., Louc¸˜a, F., 2001. As Time Goes By. Oxford University Press.Friend, A. D., White, A., 2000. Evaluation and analysis of a dynamic terrestrial ecosystem model under preindustrial conditions at the global scale.Global Biogeochemical Cycles 14 (4), 1173–1190.GEA (Ed.), 2012. Global Energy Assessment. Cambridge University Press, Cambridge, UK and New York, NY, USA and the International Institutefor Applied Systems Analysis, Laxenburg, Austria.Geels, F. W., 2002. Technological transitions as evolutionary reconfiguration processes: a multi-level perspective and a case-study. Research Policy31 (8-9), 1257 – 1274.Giraudet, L.-G., Guivarch, C., Quirion, P., 2012. Exploring the potential for energy conservation in french households through hybrid modeling.Energy Economics 34 (2), 426 – 445.Gokhale, J., Kotliko ff , L. J., 2002. Simulating the transmission of wealth inequality. The American Economic Review 92 (2), 265–269.Grubb, M., Hourcade, J.-C., Neuho ff , K., 2014. Planetary economics. Taylor Francis / Routledge, London, UK.Gr¨ubler, A., Nakicenovic, N., Victor, D., 1999. Dynamics of energy technologies and global change. Energy Policy 27 (5), 247–280.Holden, P., Edwards, N., Garthwaite, P., Fraedrich, K., Lunkeit, F., Kirk, E., Labriet, M., Kanudia, A., Babonneau, F., 2014. Plasim-entsem v1. 0:a spatio-temporal emulator of future climate change for impacts assessment. Geoscientific Model Development 7 (1), 433–451.Holden, P. B., Edwards, N. R., 2010. Dimensionally reduced emulation of an aogcm for application to integrated assessment modelling. GeophysicalResearch Letters 37 (21).Holden, P. B., Edwards, N. R., Gerten, D., Schapho ff , S., 2013. A model-based constraint on CO fertilisation. Biogeosciences 10 (1), 339–355.Holtz, G., 2011. Modelling transitions: An appraisal of experiences and suggestions for research. Environmental Innovation and Societal Transitions1 (2), 167 – 186.Holtz, G., Alkemade, F., de Haan, F., K¨ohler, J., Trutnevyte, E., Luthe, T., Halbe, J., Papachristos, G., Chappin, E., Kwakkel, J., et al., 2015.Prospects of modelling societal transitions: position paper of an emerging community. Environmental Innovation and Societal Transitions 17,41–58.Hope, C., 2011. The PAGE09 Integrated Assessment Model: A Technical Description. Cambridge judge business school working papers, Universityof Cambridge.URL IEA, 2007. Climate policy uncertainty and investment risk. Tech. rep., IEA.IEA, 2012. World Energy Outlook 2012. IEA / OECD.IEA, 2013. World Energy Outlook 2013. IEA / OECD.IEA / ETSAP, 2012. Energy technology systems analysis program.URL
IIASA, 2013. The message model.URL
IPCC, 2013a. Climate Change 2013: The Physical Science Basis. Cambridge University Press.IPCC, 2013b. Climate Change 2014: Mitigation of Climate Change. Cambridge University Press.Kahneman, D., Tversky, A., 1979. Prospect theory: an analysis of decision under risk. Econometrica 47 (2), 263–291.Kaldor, N., 1957. A model of economic-growth. Economic Journal 67 (268), 586–624.Karmeshu, Bhargava, S., Jain, V., 1985. A rationale for law of technological substitution. Regional Science and Urban Economics 15 (1), 137–141.Keen, S., 1995. Finance and economic breakdown: Modeling minsky’s ”financial instability hypothesis”. Journal of Post Keynesian Economics17 (4), pp. 607–635.Keen, S., 2011. Debunking Economics - Revised and expanded edition. Zed Books.Keynes, J. M., 1936. The General Theory of Employment, Interest and Money. Palgrave-Macmillan.Kirman, A. P., 1992. Whom or what does the representative individual represent? The Journal of Economic Perspectives, 117–136.Knobloch, F., Mercure, J.-F., 2016. The behavioural aspect of green technology investments: a general positive model in the context of heteroge-neous agents. 4CMR working papers. RL http://be.4cmr.group.cam.ac.uk/working-papers/pdf/4CMR_WP_11.pdf Labriet, M., Joshi, S., Vielle, M., Holden, P., Edwards, N., Kanudia, A., Loulou, R., Babonneau, F., 2013. Worldwide impacts of climate change onenergy for heating and cooling. Mitigation and Adaptation Strategies for Global Change, 1–26.Lee, S., Pollitt, H., Park, P.-S., 2015. Low-Carbon sustainable future in East Asia. Roudtledge.Mansfield, E., 1961. Technical change and the rate of imitation. Econometrica 29 (4), pp. 741–766.Marchetti, C., Nakicenovic, N., 1978. The dynamics of energy systems and the logistic substitution model. Tech. rep., IIASA.URL
Mas-Colell, A., Whinston, M. D., Green, J. R., 1995. Microeconomic theory. Oxford University Press.McLeay, M., Radia, A., Thomas, R., 2014. Money creation in the modern economy. Tech. rep., Bank of England.URL
McShane, B. B., Bradlow, E. T., Berger, J., 2012. Visual influence and social groups. Journal of Marketing Research 49 (6), 854–871.Meinshausen, M., Meinshausen, N., Hare, W., Raper, S., Frieler, K., Knutti, R., Frame, D., Allen, M., 2009. Greenhouse-gas emission targets forlimiting global warming to 2c. Nature 458 (7242), 1158–1162.Mercure, J.-F., 2012. Ftt:power : A global model of the power sector with induced technological change and natural resource depletion. EnergyPolicy 48 (0), 799 – 811.URL http://dx.doi.org/10.1016/j.enpol.2012.06.025
Mercure, J.-F., 2015. An age structured demographic theory of technological change. J. Evol. Econ. 25, 787–820.URL http://arxiv.org/abs/1304.3602
Mercure, J.-F., Lam, A., 2015. The e ff ectiveness of policy on consumer choices for private road passenger transport emissions reductions in sixmajor economies. Environ. Res. Lett. 10 (064008).Mercure, J.-F., Pollitt, H., Chewpreecha, U., Salas, P., Foley, A., Holden, P., Edwards, N., 2014. The dynamics of technology di ff usion and theimpacts of climate policy instruments in the decarbonisation of the global electricity sector. Energy Policy 73 (0), 686 – 700.Mercure, J.-F., Salas, P., Foley, A., Chewpreecha, U., Pollitt, H., Holden, P. B., Edwards, N. R., 2015. Complexity, economic science and possibleeconomic benefits of climate change mitigation policy. Under review, preprint available on ArXiv.URL http://arxiv.org/abs/1310.4403 Meyer, B., Lutz, C., 2007. The ginfors model: Model overview and evaluation. Tech. rep.URL
Meyer, M., Distelkamp, M., Ahlert, G., Meyer, B., 2013. Macroeconomic modelling of the global economy-energy-environment nexus: Anoverview of recent advancements of the dynamic simulation model ginfors. Gws discussion paper 2013 /
5, GWS.URL http://papers.gws-os.com/gws-paper13-5.pdf
Miles, D., 1999. Modelling the impact of demographic change upon the economy. The Economic Journal 109 (452), 1–36.Morton, D. C., DeFries, R. S., Shimabukuro, Y. E., Anderson, L. O., Arai, E., del Bon Espirito-Santo, F., Freitas, R., Morisette, J., 2006. Croplandexpansion changes deforestation dynamics in the southern brazilian amazon. Proceedings of the National Academy of Sciences 103 (39),14637–14641.NIES, 2012. Asia-pacific integrated model.URL
Nordhaus, W. D., 2010. Economic aspects of global warming in a post-Copenhagen environment. Proceedings of the National Academy of Sciences.Nordhaus, W. D., 2013. Dice-2013r model.URL http://aida.wss.yale.edu/~nordhaus/homepage/Web-DICE-2013-April.htm
OXFAM, 2012. A safe and just space for humanity. Tech. rep., OXFAM.URL
Oyebamiji, O. K., Edwards, N. R., Holden, P. B., Garthwaite, P. H., Schapho ff , S., Gerten, D., 2015. Emulating global climate change impacts oncrop yields. Statistical Modelling.Perez, C., 2001. Technological Revolutions and Financial Capital. Edward Elgar.Piesse, J., Thirtle, C., 2009. Three bubbles and a panic: An explanatory review of recent food commodity price events. Food Policy 34 (2), 119 –129.PIK, 2011. Dice-2013r model.URL Pollitt, H., Mercure, J.-F., 2015. The role of money and the financial sector in energy-economy models used for assessing climate policy. C-EENRGworking papers.URL