Agent-Based Macroeconomics

Frontier If a macroeconomic model assembled one agent at a time out of administrative data can forecast an economy at least as well as the equilibrium and reduced-form models central banks already run — which one published comparison shows for one country, and which 7 says has never been tested under conditions capable of settling it — then a single object answers the aggregate question and the distributional question in the same run. The Austrian model at the centre of this brief runs at one agent per person, about 8.86 million of them against a country of roughly 8.8 million inhabitants, across 64 industry sectors. The standard comparison model, of the Smets and Wouters type, carries seven observed variables and contains no sectors at all.

Established The prize statable in units is coverage per model run: 64 sector-level forecasts and a household-level income distribution fall out of the same simulation that produces the GDP number, where the incumbent buys sectors and distributions from separate satellite models reconciled by judgement. That reconciliation is a named institutional problem. The Bank of Canada published two staff discussion papers in 2023, one a blueprint for a fourth generation of its projection models and one on bringing research models into projections, and a 2025 Bank of England survey records it as the first central bank to give this model class a formal place in its official modelling strategy.

Frontier The gain accrues first to the few countries whose statistical systems can produce a one-to-one synthetic population, and to model vendors second. Austria, Canada, Hungary and the euro area recur in this literature because they hold linked national accounts, input-output tables, census and business demography data of the kind the Austrian model consumes. A country without that pipeline cannot build the model at all. The Oxford group behind the 38-country extension lists a simulation company among its collaborators and funders, and its senior author’s 2025 journal affiliation includes a modelling firm. Neither fact makes a result wrong; both mean the enthusiasm has owners.

Speculative The condition that caps the case is that nobody has run the comparison the case needs. The Austrian result is one published comparison, on one country, over 39 reference quarters, against benchmarks its own authors selected and implemented. The 38-country follow-up is scored against an AR(1) and against the earlier agent-based model, not against the DSGE and Bayesian VAR systems central banks actually run. Whether behavioural realism buys forecast accuracy, or merely buys fit, is the quantity 5 says nothing currently measures.

Established The object of this brief is a modelling paradigm, not a policy or a device. Two ways of writing down an economy are in competition. The incumbent is the dynamic stochastic general equilibrium model: a few optimising agents, usually collapsed to one representative household and one firm, forming rational expectations, clearing markets in equilibrium, linearised around a steady state, driven by exogenous shocks, with parameters estimated by Bayesian methods on a handful of aggregate series. The challenger is the agent-based model: many heterogeneous agents following explicit behavioural rules, interacting directly and through markets, with aggregates emerging from simulation rather than imposed.

Established The difference that matters for evidence is where the discipline comes from. A DSGE model is disciplined by internal consistency and has few free parameters relative to the data it is fitted to. An agent-based model is disciplined, when at all, by external data: rules read off surveys and microdata, initial conditions read off national accounts. The two fail in opposite directions — the first can be internally immaculate and empirically empty, the second can fit everything and identify nothing — and 15 takes each failure in its strongest form.

Established The Austrian model is the concrete instance this brief is tethered to. Poledna, Miess, Hommes and Rabitsch, in the European Economic Review in January 2023, open their abstract: “We develop the first agent-based model (ABM) that can compete with benchmark VAR and DSGE models in out-of-sample forecasting of macro variables.” It uses national accounts, sector accounts, input-output tables, government statistics, and census and business demography data; it represents all activities classified by the European System of Accounts (ESA 2010); and it populates every sector at a scale of one to one between model and data, each agent standing for a natural or legal person.

Frontier What is new in the last five years is not the idea but the arithmetic. Agent-based macroeconomics existed for three decades as a qualitative programme, calibrated to reproduce stylised facts and judged on shapes. The 2024 Oxford extension states the transition in its own abstract: “In the last few years, economic agent-based models have made the transition from qualitative models calibrated to match stylised facts to quantitative models for time series forecasting, and in some cases, their predictions have performed as well or better than those of standard models”. That sentence is the claim under test, and “in some cases” is doing more work than it lets on.

Established The tradition is not one thing, and conflating its branches inflates its record. The Schumpeter-meeting-Keynes family of Dosi and Roventini, published as a Cambridge Element in 2025, is an evolutionary programme aimed at innovation, industrial and fiscal policy, with agents following “quite simple, empirically based, behavioural rules” in an economy treated as “a complex evolving system”. The financial-stability branch inside central banks models housing markets and interbank networks, not the macroeconomy. Only the forecasting branch has entered a like-for-like competition with DSGE models on their own ground, and it is the branch assessed here.

Established The anchor result is real, published, and narrower than its reputation. Poledna, Miess, Hommes and Rabitsch (European Economic Review 151, article 104306, online October 2022) compared an agent-based model of Austria against VAR, AR, vector error correction and DSGE benchmarks on root mean squared error, the DSGE being Smets and Wouters adapted to Austria by Breuss and Rabitsch. Parameters were calibrated to 39 different reference quarters from 2010Q1 to 2019Q3; for each, 500 Monte Carlo simulations were run and the model solved recursively for twelve periods, the maximum horizon.

Established The commission this brief was written against described the anchor as beating the official forecasting institutes, and that is not what the paper claims. The published abstract names VAR and DSGE benchmarks. The working-paper version, read here in full, compares against VAR, AR and DSGE and contains no comparison with the Oesterreichische Nationalbank, WIFO or the Institute for Advanced Studies. The distinction is not pedantry: institutional forecasts carry judgement, off-model information and a release calendar, and beating them is a harder claim. The commission also named J. Doyne Farmer among the authors; he is not one. He is senior author of the 38-country extension below, a different result with different properties.

Established What the comparison found is a tie at short horizons and an advantage at three years. In the working paper’s words the model delivers similar forecast performance to the VAR for horizons up to two years and significantly improves on VAR forecasts for longer horizons up to three years; it delivers similar performance to the standard DSGE model, and both improve on the AR model for household consumption and investment. The scale is one to one: 4,729,215 economically active and 4,130,385 economically inactive persons, against a country the paper describes as having about 8.8 million inhabitants, across 64 industry sectors.

Frontier The statistical weight of the three-year result is far lower than “significantly” suggests, and this is the most important caveat in the brief. Thirty-nine quarterly origins spanning 2010Q1 to 2019Q3 yield, at a twelve-quarter horizon, roughly three non-overlapping three-year forecasts. Overlapping horizons share data and shocks, so the effective sample where the advantage appears is small, and the window is one business-cycle phase: a recovery from the euro-area crisis, with no recession and no inflation surge in it.

Frontier The 38-country extension is the strongest quantitative evidence the paradigm has, and its benchmarks are weaker than the Austrian paper’s. Wiese, Kaszowska-Mojsa, Dyer, Morán, Pangallo, Lafond, Muellbauer, Calinescu and Farmer (Oxford INET working paper 2024-06, 27 September 2024) add housing markets, synthetic populations with income, wealth and consumption heterogeneity, richer behavioural rules and an enhanced credit market, and calibrate for all 38 OECD members using neural posterior and neural ratio estimation on the first moments of five time series between 1990Q1 and 2013Q1. Against an AR(1), reported relative improvement in median root mean squared error is 41 per cent for GDP at one quarter rising to 62 per cent at three years, and 44 to 63 per cent for inflation, with lower error than the benchmarks for all considered aggregates in 33 of 38 countries.

Frontier The margin over the previous agent-based model, not over the AR(1), is the number that matters, and it is modest. Against the Austrian model, 65 per cent of estimated Bayes factors support the new model. That is a majority, not a rout, and it is a Bayesian model comparison rather than a test of predictive accuracy. No DSGE model appears in the comparison at all. Beating an AR(1) by 40 to 60 per cent and a predecessor on 65 per cent of Bayes factors is compatible with the paradigm being decisively better than time-series benchmarks and no better than a well-specified Bayesian VAR, and the paper cannot distinguish those cases.

Established On the other side, the DSGE forecasting record is not the failure the popular account describes, and not a success either. Edge and Gürkaynak evaluated the Smets and Wouters model’s real-time forecasts against Bayesian VAR, Federal Reserve Greenbook, Blue Chip and random-walk benchmarks over 1992 to 2004 and 1992 to 2009. They found the DSGE forecasts had the lowest root mean squared error for both inflation and growth at longer horizons, then wrote that the DSGE model in an absolute sense did a very poor job of forecasting, that all three methods are poor, and that combining them does not lead to much improvement either. Winning a comparison in which everything is bad establishes little.

Established The euro-area and Swedish evidence is a split decision by horizon, which is the same pattern the Austrian paper reports for its own model. Christoffel, Coenen and Warne (ECB working paper 1185, 2010) found the New Area-Wide Model competitive over 1999Q1 to 2006Q4 on real GDP, trade and employment, weakest on nominal wage growth — a steady-state 3.1 per cent against outturns near 2.3 per cent produced systematic overprediction — while a large Bayesian VAR with a mixed prior tended to outperform at short horizons. Adolfson, Andersson, Lindé, Villani and Vredin (International Journal of Central Banking, 2007) found the Riksbank’s DSGE and official forecasts about equally precise for inflation to a year, the DSGE better at five to eight quarters, the BVAR strong to six, and the combination better than any of them.

Established That paper also names the flaw that makes nearly every comparison in this literature uninterpretable. Adolfson and co-authors write that the formal models possibly have an information advantage relative to the official forecasts, since the BVAR and DSGE forecasts are based on ex post data on GDP, and note that preliminary GDP data arrive about a quarter late and are subsequently revised. A model scored on revised data is answering an easier question than a forecaster faced. Warne’s ECB working paper 2768 is one of the few that runs on real-time vintages, 2001Q1 to 2019Q4, for a Galí-Smets-Wouters euro-area model.

Established Warne’s result is the cleanest demonstration that in-sample fit and out-of-sample accuracy point in opposite directions. The adaptive-learning variant has a log marginal likelihood roughly 22 log-units above the rational-expectations variant over the full estimation sample — overwhelming by any conventional reading — and the rational-expectations variant nevertheless predicts real GDP growth more accurately out of sample, with density scoring rules agreeing. Any argument that a model should be preferred because it fits better, or because its rules are more realistic, has to survive this.

Established Central banks have adopted agent-based models, but mostly not for macroeconomic forecasting. Bank of England staff working paper 1122 (February 2025), by Borsos, Carro, Glielmo, Hinterschweiger, Kaszowska-Mojsa and Uluc, surveys over twenty central banks and several other public institutions. The dominant use is financial stability: interbank contagion at the ECB and Banco de España, housing-market models at the Bank of England and the Magyar Nemzeti Bank, whose results have appeared in its Financial Stability Reports since 2016. That is a substantial record and it is not evidence about macroeconomic forecasting.

Established The exception is the Bank of Canada, and it is the most consequential institutional fact here. The same survey describes it as a pioneer in incorporating these models into its general toolkit for macroeconomic projections and policy analysis, and as the first central bank to formally acknowledge a role for the approach within its official modelling strategy. The vehicle is CANVAS — Hommes, He, Poledna, Siqueira and Zhang, Journal of Economic Dynamics and Control, 2024 — which recalibrates the Austrian model to Canada and adds greater household heterogeneity, a dynamic monetary policy rule and a more sophisticated price-setting rule, and is reported as particularly informative during the COVID-19 lockdowns and the supply-chain disruption that followed.

Frontier “Particularly informative” is a judgement by the institution that built it, and no scored record of CANVAS forecasts against the Bank of Canada’s own projection models has been published that this brief could find. That is the shape of the whole evidence base: the strongest institutional endorsement is an internal assessment, and the strongest quantitative evidence is a comparison run by a model’s authors against benchmarks they chose. Neither is dishonest. Neither would settle the question.

Established Where the paradigm has a genuine externally checkable forecasting record is the pandemic. Pichler, Pangallo, del Rio-Chanona, Lafond and Farmer built a dynamic disequilibrium input-output model that made real-time forecasts for the decline of sectoral and aggregate economic activity in the United Kingdom in the second quarter of 2020 (Journal of Economic Dynamics and Control 144, November 2022). Pangallo and twelve co-authors, in Nature Human Behaviour in 2024, coupled epidemic and economic modules over 416,442 synthetic individuals representative of the New York metropolitan area, daily from 12 February to 30 June 2020, reporting Pearson correlations of 0.82 for industry-level employment change and 0.75 for contact-pattern reductions.

Established The methodological literature has moved fast enough that the old objection — that these models cannot be estimated, only calibrated by hand — is no longer simply true. Grazzini, Richiardi and Tsionas set out Bayesian estimation for agent-based models in 2017, covering kernel-density likelihoods with Markov chain Monte Carlo, parametric approximations and approximate Bayesian computation. Delli Gatti and Grazzini added a procedure respecting the time structure of simulated and observed series, plus a forecasting method, in 2020. Martinoli, Moneta and Pallante proposed calibration and validation by statistical causal search, comparing structural VARs identified by independent component analysis on simulated against observed data. The 38-country paper uses neural simulation-based inference, and a related Oxford method paper took the best-paper award at the 2024 Conference on Uncertainty in Artificial Intelligence.

Frontier The gap between those methods and the models they must estimate is roughly an order of magnitude in parameter count, and nobody has closed it. The published Bayesian demonstrations run from a price-discovery model with one parameter to a macroeconomic model with nine unknown parameters. The models in the forecasting competition carry behavioural rules for consumption, pricing, hiring, investment, credit and housing across dozens of sectors. Calibrating to the first moments of five time series is a choice forced by computation, not a statement that five moments identify the model.

Frontier Does behavioural realism buy forecast accuracy, or does it buy fit? This is the load-bearing question and everything else is downstream. Warne’s adaptive-learning result shows the two can move in opposite directions by a wide margin inside one model family. No published study isolates the contribution of agent heterogeneity to forecast accuracy by removing it and re-scoring, which is the obvious test and is set out in 7.

Frontier Is the long-horizon advantage real or an artefact of overlapping windows? The Austrian edge appears at three years, where 39 origins give roughly three independent observations. The 38-country paper’s largest margins over AR(1) are also at three years, which is where an AR(1) is weakest by construction because it reverts to an estimated mean. A structural model beating a mean-reverting univariate benchmark at long horizons is close to definitional.

Frontier Are the behavioural rules identified? The DSGE identification problem is documented and quantified. Romer put it arithmetically: a model with m variables has m-squared parameters and m equations, so Smets and Wouters with seven variables has 49 parameters and seven equations, leaving 42 quantities supplied from outside the time series, and adding expectations doubles the count to 91. The agent-based analogue has never been stated with that precision, because nobody has written down an order condition for a model whose mapping from parameters to moments exists only by simulation.

Speculative If two materially different rule sets produce indistinguishable fit, the model cannot support the policy counterfactuals that are its main selling point. This is equifinality, and in a simulation model it is the default expectation rather than a worry. The validation literature — Fagiolo, Guerini, Lamperti, Moneta and Roventini organise it around comparison of artificial with real data, calibration and estimation, and parameter-space exploration — treats it as open, and Tieleman’s 2021 proposal is explicitly a methodology for validating mechanisms separately from targets.

Frontier Does the Lucas critique bite agent-based models? The standard claim is that it does not: rules estimated from microdata are policy-invariant where reduced-form aggregate relations are not. The steelman of the other side is that a rule fitted within one policy regime is a reduced form over that regime’s expectations too, so a fixed-rule simulation misforecasts a regime change for exactly the reason Lucas gave in 1976. This has never been tested.

Frontier Is the incumbent’s diagnosis of its own problems the right one? Blanchard’s 2016 policy brief concluded that current DSGE models are seriously flawed but eminently improvable and central to the future of macroeconomics, and should become less insular and less imperialistic and share the scene with other types of general equilibrium model. Galí located the flaw in two assumptions: the infinitely-lived representative household, which rules out persistent negative natural rates and rational bubbles, and the restriction to stationary linear fluctuations, which rules out the nonlinear boom-bust dynamics and credit freezes that make financial crises interesting. Both are reform programmes, not replacement programmes.

Frontier The rise of heterogeneous-agent DSGE models is the most serious threat to the agent-based case and is usually omitted from it. If the distributional content agent-based models sell can be had inside an optimising framework with a tractable estimation problem, the marginal value of abandoning optimisation is whatever remains: network structure, disequilibrium dynamics, non-clearing markets. Galí endorses the heterogeneous-agent turn while warning that computational burden limits adoption, which is the same constraint that binds the challenger.

Speculative Nobody knows what the right number of agents is. The Austrian model runs at one to one with the population; the New York pandemic model runs 416,442 synthetic individuals for a metropolitan area many times larger; the Schumpeter-meeting-Keynes family runs a few thousand. No published study establishes where forecast accuracy stops improving with agent count, so the field cannot say whether one-to-one scale is a scientific requirement or an aesthetic commitment with a large compute bill attached.

Handwave The claim that complexity is why economics failed in 2008, and that agent-based models are the remedy, is an argument by analogy never cashed out. Farmer and Foley made the case in Nature in 2009; Haldane and Turrell restated it in the Oxford Review of Economic Policy in 2018, arguing that agent-based models suit domains where complexity, heterogeneity, networks and heuristics matter. The step from “the economy is a complex system” to “therefore this class of simulation forecasts it better” is where the argument works by assertion, and the 2023 and 2024 papers exist because that step needed evidence.

Established Compute is the first hard constraint and it scales multiplicatively with everything a serious evaluation requires. The Austrian exercise runs 500 Monte Carlo simulations per reference quarter across 39 quarters, each solved twelve periods forward. Multiply by 38 countries, then by the candidate parameter vectors a Bayesian estimator must evaluate, and the cost of one honest calibration is why the 38-country paper calibrates to five moments rather than a full likelihood. The Bank of England survey lists computational intensity among its named difficulties, alongside the complexity of defining behavioural rules.

Established Estimation methods exist but have been demonstrated at a parameter count far below the models in use. Nine unknown parameters in the published Bayesian demonstration, against dozens of behavioural rules in a forecasting model, is not a gap better sampling closes on its own. It is the difference between a problem known to be identified and one not known to be.

Frontier Data access, not data existence, is the binding constraint on building the model outside a handful of countries. The Austrian model consumes national accounts, sector accounts, input-output tables, government statistics, census and business demography data linked at person and firm level. In most jurisdictions those files exist behind separate legal regimes with separate custodians. This is the same obstacle Economic Resilience reports for deep-tier supply chain data, and the remedy is the same kind of object: an access rule, not a discovery.

Frontier Real-time data vintages are the bottleneck nobody in the agent-based literature has paid for. A model calibrated on revised national accounts and scored on revised national accounts answers a question no policymaker was asked. The DSGE literature built real-time infrastructure decades ago: Edge, Kiley and Laforte evaluated over 70 forecasts presented to the FOMC between 1996 and 2005 on a real-time basis, and Warne’s euro-area evaluation runs on vintages from 2001Q1 to 2019Q4. This brief found no agent-based macroeconomic forecasting evaluation conducted on real-time vintages.

Established Traceability is a bottleneck on use rather than construction, and it is the one central banks report hurting most. The Bank of England survey states that because agent interactions are nonlinear, tracing the pathways from specific behaviours or rules to emergent outcomes demands significant time and effort. A projection round has a deadline. A model whose answer cannot be explained to a governance committee inside that deadline does not enter the round, whatever its root mean squared error.

Frontier There is no agreed scoring standard, and the absence is load-bearing. The Austrian paper scores root mean squared error; the Oxford extension reports Bayes factors; Warne reports point forecasts and density scoring rules. Three currencies for one question is how a literature accumulates results that cannot be added up, and it is the first thing 7 has to fix.

Frontier The absence of a pre-registration norm in macroeconomic model comparison is a bottleneck disguised as a culture. Every comparison cited here was designed after the data existed, by authors who knew how their model performed, choosing which benchmarks to implement and which variables to report. None of that is misconduct and all of it is degrees of freedom. Economics has a working pre-registration apparatus for randomised trials and none for model competitions.

Speculative The deepest bottleneck may be that the quantity in dispute is not forecast accuracy at all. If both paradigms forecast badly in absolute terms, as Edge and Gürkaynak concluded for the DSGE family, a tournament settles a contest neither side much cares about, and the real question is which model gives better policy counterfactuals — for which there is no out-of-sample test, because the counterfactual is unobserved by definition.

Established The paradigm depends on simulation-based inference improving faster than model complexity grows. Neural posterior estimation, neural ratio estimation and approximate Bayesian computation with path signatures are the current front. The Oxford group’s standing project on estimating parameters and initial conditions exists because, in its own framing, many agent-based models — particularly large-scale macroeconomic variants — are not formulated in terms of smooth functions, which is what makes classical estimation inapplicable.

Frontier It depends on a validation criterion that is not moment-matching. Matching first moments of five series is a weak test that many wrong models pass. The causal-search approach — estimate a structural VAR with independent component analysis on simulated and on observed data, then compare the recovered causal structures — is the most promising published alternative because it asks whether the model has the right mechanism rather than the right averages.

Established It depends on real-time data infrastructure that already exists on the incumbent’s side. Vintage databases, dated forecast archives and a release calendar are what make a forecast comparison interpretable, and the DSGE literature has used them since the 1990s. Nothing has to be invented. It has to be adopted.

Frontier It depends on microdata on behaviour, not just on stocks and flows. Reading behavioural rules off surveys requires surveys that measure what agents do when conditions change: expectations, search, price-setting frequency, consumption response to transfers. Where these exist they are national and idiosyncratic, which is why the rules in the models cited here are largely inherited from the modelling tradition rather than estimated afresh per country.

Speculative It depends on an identification result nobody is currently producing. The question is whether, for a stated class of agent-based macro model, the mapping from behavioural parameters to observable moments is injective on the relevant region. That is a statement about a simulator, answerable at least locally by numerical experiment, and no paper this brief read attempts it at forecasting scale.

Speculative It may depend on abandoning the one-to-one scale that gives the approach its rhetorical force. If forecast accuracy saturates at a hundred thousand agents, the one-to-one model is paying a large compute bill for a property with no measured payoff, and the honest version of the programme is smaller, cheaper and much easier to estimate.

Frontier The decisive experiment is a multi-year pre-registered forecasting tournament: agent-based, DSGE, Bayesian VAR and a naive benchmark, one country, one release calendar, scored on real-time data vintages against criteria fixed before the first forecast is made. Nothing would change this brief’s assessment more. Every comparison in the literature fails at least one of those conditions, and most fail three.

Established Every component of the tournament already exists. The models are published: the Austrian and Canadian agent-based models, the Smets and Wouters family and its national adaptations, Bayesian VARs with standard priors, and an AR(1) as the null. Real-time vintage databases have been used in DSGE evaluation since the 1990s. The scoring statistics are standard. The design cost is a protocol document; the running cost is compute and a custodian.

Frontier The pre-registration has to bind the things that are currently free. Before the first origin: the model code and its parameter vector deposited and hashed with a third party; the variable list; horizons of one to twelve quarters, matching the Austrian model’s own maximum; the loss functions, point and density; the test of equal predictive accuracy; the treatment of revisions, meaning which vintage counts as the outturn; and the rule for a model that fails to converge. Re-estimation on each new vintage is allowed and specified in advance; re-specification is not.

Frontier Run length decides whether the result means anything, and the arithmetic is unforgiving. Five years of quarterly origins gives twenty origins and roughly five non-overlapping three-year forecasts; ten years gives forty and roughly ten. The Austrian window contained no recession. A tournament that cannot span at least one turning point will settle nothing about behaviour away from the sample mean.

Established The choice of country is not neutral and Canada is the obvious site. It is the only jurisdiction where a central bank has given an agent-based model a formal place in its official modelling strategy, so both entrants have institutional owners inside one building and neither is a straw man. Austria is the alternative, because the original result is there and a replication would test it directly.

Speculative The second experiment is an ablation, and it is cheaper than the tournament by orders of magnitude. Take the Austrian or Canadian model as built; produce variants with heterogeneity progressively removed — collapse households to deciles, then to one; collapse 64 sectors to ten, then to one; replace the network of firm-to-firm links with an aggregate input coefficient — and re-score each variant on the same origins. If accuracy is flat across that sequence, the one-to-one scale is decoration. If it degrades monotonically, the field has its first measurement of what heterogeneity is worth in error units. No such curve has been published.

Speculative The third is an identification probe on the simulator itself. Fix a model, sample widely separated points in behavioural-parameter space, simulate, and ask whether the calibration target — the five moments the 38-country paper uses, or any proposed alternative — can tell them apart. This has no data requirement, is answerable in weeks of compute, and would convert the equifinality objection from a rhetorical complaint into a measured property of a named model.

Frontier The fourth is the regime-change test, the only direct test of the Lucas critique claim. Calibrate a model on a window containing one policy regime, then ask it to forecast across a documented regime change it never saw — an inflation-targeting adoption, a fiscal rule, a large transfer programme — against a reduced-form benchmark calibrated on the same window. The agent-based case predicts the structural model degrades less. Nobody has run it, and it needs no new model.

Frontier The fifth is an independent replication of the Austrian result from released code. A team with no stake re-runs the 39 reference quarters, the 500 Monte Carlo draws per quarter and the twelve-period horizons, and reports whether the ranking against VAR, AR, VECM and DSGE reproduces. This is the cheapest test in the brief and the one whose failure would be most informative, because a forecasting claim that cannot be reproduced from its own code is not a claim about the world.

Speculative The sixth is an ensemble test rather than a contest. Compute the correlation between agent-based and DSGE forecast errors on the same origins, and score the optimal combination. Forecast combination beat both components in the Riksbank comparison; if it does so here, the policy conclusion is to fund a third model rather than replace the first, and the rivalry framing organising this literature is answered by arithmetic rather than argument.

Established The natural experiment already ran once, and the field did not score it. COVID-19 hit with agent-based, DSGE and reduced-form models all in the field, at a moment when the shock was observable in advance and its sectoral incidence was in the policy order. The pandemic papers report their own accuracy, and the Bank of Canada reports CANVAS as particularly informative, but no independent party collected the dated forecasts of all three model classes and scored them on a common criterion. That archive may still be recoverable from institutional records, and recovering it is the fastest route to a partial answer.

Frontier What would move this brief toward the agent-based paradigm is a tournament win at short horizons. The long-horizon advantage is the one an AR(1) is designed to concede. A model that beats a Bayesian VAR at one to four quarters on real-time vintages, pre-registered, would be a genuinely new fact about macroeconomics.

Frontier What would move it the other way is a null on the ablation. If collapsing millions of agents to a few dozen types costs nothing measurable in forecast error, the empirical case reduces to the sector-level and distributional outputs that nothing currently scores, and the honest recommendation becomes a much smaller and cheaper model.

Handwave The claim that no tournament is needed because policy relevance matters rather than forecasting is available to both sides and should be refused from both. It is true that a model can be useful without forecasting well. It is also the argument that makes a modelling paradigm unfalsifiable, and both camps reach for it when their forecasts lose. If neither side will be bound by a scored comparison, the disagreement is not scientific and should be described as something else.

Established Building one of these models is a data engineering project before it is an economics project. The Austrian inputs are national accounts, sector accounts, input-output tables, government statistics, census and business demography data, reconciled to the European System of Accounts classification and to each other, at one agent per natural or legal person. The reconciliation is the work: national accounts and business registers disagree, and a one-to-one model has nowhere to hide a residual.

Frontier Synthetic population generation is the reusable component and the one most likely to be standardised. The New York pandemic model built 416,442 individuals from mobility data, the American Community Survey, Bureau of Labor Statistics data, input-output tables and census microdata; the 38-country model builds populations carrying income, wealth and consumption heterogeneity for every OECD member. This is the piece a statistical agency could supply as a product, and the piece whose absence stops a country entering the field at all.

Frontier The platform is the deliverable the field is missing. The 38-country paper reports building one so alternative models can be built, run and evaluated, and hopes this encourages future work. Until such a platform is used by a team that did not write it, the field has a strong claim and no independent test of it — the same structural problem Forecasting Replication documents for published findings generally.

Frontier A tournament needs one piece of engineering that does not exist: a custodian system that accepts a hashed model and a dated forecast and releases neither until the outturn. This is modest software with an institutional wrapper, and it is the same instrument Decision Markets in Governance and Technology Forecasting both depend on for their own scoring claims.

Speculative Interpretability tooling is the engineering gap with the clearest payoff. If tracing an emergent outcome back to the rules that produced it demands significant time and effort, then the tool that automates the trace — attribution of an aggregate movement to agent classes and rules — is what converts the model from a research object into a committee-usable one. No such tool is standard.

Frontier The distributional outputs need an evaluation harness of their own. Forecasting the income distribution has no incumbent benchmark and no agreed loss function, so the paradigm’s most distinctive product is unscoreable — a gap that also runs through Future Labour Markets and Future Taxation Models, where distributional projections are used without anyone reporting how accurate the last ones were.

Established Epidemic modelling supplied the method and the credibility. The coupled epidemic-economic model of the New York metropolitan area is an agent-based epidemiological model with an economy attached, and the pandemic is where this class of model earned its central-bank hearing.

Established Simulation-based inference is a machine learning field that has quietly become macroeconomics’ dependency. Neural posterior estimation, neural ratio estimation and approximate Bayesian computation are developed in statistics and machine learning venues, and the best-paper award at the 2024 Conference on Uncertainty in Artificial Intelligence went to work on approximate Bayesian computation with path signatures done by an economics group with a simulation company. If agent-based macroeconomics succeeds, it will be because that field solved its estimation problem.

Frontier Multi-agent artificial intelligence is adjacent in method and opposite in purpose. The techniques for training and analysing populations of interacting artificial agents overlap almost completely with the machinery here, and Multi-Agent Intelligence Systems owns that question. The difference is that an economic model is judged against an observed economy and a multi-agent system against a task, and the second has no equivalent of the identification problem in 4.

Frontier Financial-stability modelling is the branch already operational and worth separating from the forecasting branch. Interbank contagion networks and housing-market models are in policy use at several central banks and appear in published financial stability reports; they are agent-based, validated against different data, and their success does not transfer to the macroeconomic forecasting claim. Future Capital Markets carries the market-structure side.

Established Input-output and production-network economics is the part of the mainstream the programme has already been absorbed into. The pandemic forecasting success ran through a dynamic disequilibrium input-output model rather than a full agent-based one, and the sectoral detail in the Austrian model is input-output structure with agents attached. The strongest version of the claim may be that networks matter, which the mainstream can and partly has adopted without abandoning optimisation — the same propagation methods Infrastructure Resilience uses for physical systems.

Established The institutional fact that matters most is that one central bank has committed and the rest have not. The Bank of Canada is recorded as the first to formally acknowledge a role for agent-based modelling within its official modelling strategy, and as a pioneer in incorporating it into the general toolkit for macroeconomic projections and policy analysis. Elsewhere the models sit in financial stability divisions, informing reports rather than projections.

Frontier One committed institution is enough to make a tournament possible and not enough to make it credible. If the Bank of Canada both hosts the tournament and enters it, the result is an internal assessment with a larger budget. The design needs a custodian that is not a competitor: a statistical agency, a research council, or a consortium whose scoring party has no model in the race.

Established Nobody currently pays for being scored, which is the whole institutional problem in one sentence. A central bank publishes a projection and is judged on the economy; a modelling group publishes a comparison it designed and is judged on citations. Neither rewards depositing a dated forecast with a third party and living with the result, which is why Economic Resilience reaches the same conclusion from the opposite direction: its cheapest high-value experiment is also a pre-registration nobody has run.

Frontier The funding structure of the agent-based programme is unusual for macroeconomics and should be stated plainly. The Oxford parameter-estimation programme lists an asset manager, a national energy scholarship scheme and an engineering doctoral training centre among its funders, with a simulation company as collaborator; the senior author’s 2025 journal affiliation includes a modelling firm. Private funding is how a method gets built when no ministry will pay, but the field’s most active group has a commercial interest in the answer, and a tournament is the instrument that makes that irrelevant.

Established The incumbent has an institutional advantage that has nothing to do with being right. DSGE models are taught, their estimation packages are standard, their staff are trained in graduate programmes that produce hundreds a year, and their outputs fit the committee formats central banks already use. The Bank of England survey’s list of practical difficulties — specialised skills, interdisciplinary teams, hard-to-trace causal pathways — describes why a better method can still lose.

Frontier Data custodianship is the other lever. A statistical agency publishing a maintained synthetic population under a research licence would remove the largest fixed cost of entry and make independent replication possible for the first time. That is a decision about access rules, not a research result, and it is the kind of thing Scientific Funding Models treats as the actual binding constraint on a field’s progress.

Handwave The claim that the profession will converge once the evidence is in assumes the evidence will be produced, and the argument for that is circular. The evidence requires a tournament; the tournament requires an institution to fund being scored; the funders take their cue from the profession. This is a coordination failure with a named solution and no named owner.

Established The choice of modelling paradigm is a distributive choice made before any policy question is asked. A representative-agent model has no distribution in it, so distributional consequences are invisible to it by construction and must be supplied by a separate exercise. An agent-based model has a distribution whether or not anyone asks. Which model an institution runs decides which harms are countable, and that decision is usually made on methodological grounds by people who would not call it an ethical one.

Established The distributional content is not hypothetical: it is what the models are being used to say. The New York pandemic study concludes that low-income workers in in-person, customer-facing occupations face the strongest health-economy trade-off, and reproduces the pattern that low-income individuals were more likely to become unemployed while reducing consumption less than high-income individuals — a result it reports as emerging without being targeted in calibration. The Canadian model is used on the uneven impact of lockdown across households and industries. These are claims the incumbent cannot make at all.

Frontier A model of every person raises a privacy question that synthetic-population construction only partly answers. The agents are synthetic, matched to marginal and joint distributions rather than to individuals, which is the standard and correct protection. The New York model nonetheless consumed privacy-preserving GPS mobility data from a commercial vendor combined with venue information, and re-identification risk in such combinations is a live research question rather than a settled one.

Speculative False precision is the specific harm this paradigm risks, and it mirrors the incumbent’s. A model reporting an outcome for every household invites reading household-level output as a prediction about households, when the calibration targets were five aggregate moments. The discipline that would prevent this — reporting which outputs are calibrated, which validated and which merely emitted — is not standard practice in the published papers.

Speculative The strongest ethical argument for the paradigm is that it makes population coverage an explicit choice. Someone has to decide whether the synthetic population includes undocumented workers, informal employment or the institutionalised. In an aggregate model that decision is made silently by the data source; in a one-to-one model somebody has to write the line. Making an omission visible is not the same as fixing it, and it is better than not knowing.

Frontier The civilizational question is not whether these models forecast better but whether a society can be governed by a model it cannot reduce to an argument. A DSGE model can be written as equations and debated by people who read equations. A simulation with millions of interacting agents can be debated only through its outputs and its author’s account of it. If policy shifts toward the second, the number of people who can contest a macroeconomic claim on its merits falls sharply.

Speculative Modelling capacity becomes a form of sovereignty if the method wins. A one-to-one model requires linked administrative microdata that only some states hold and only some can legally combine. A world in which credible macroeconomic models exist for Austria, Canada, Hungary and the OECD members but not elsewhere is one where policy advice to everyone else is produced abroad, on foreign calibrations, by whoever owns the platform.

Frontier Monoculture is the failure mode a tournament could cause rather than prevent. If one paradigm is declared the winner and adopted widely, correlated model error becomes systemic: every institution misses the same turning point in the same way, which is a fair description of 2008. The finding that forecast combination beat both components in the Riksbank comparison points toward deliberately maintained diversity, which is an argument against the winner-takes-all framing this brief’s own decisive experiment invites.

Handwave The claim that better macroeconomic models would have prevented the last crisis, and will prevent the next, is the field’s founding story and has no evidence behind it. The 2008 failure was as much about incentives, data on off-balance-sheet exposure and regulatory attention as about model class, and no counterfactual exercise has shown that a model of the available data would have produced a different decision. The argument is repeated because it is motivating, not because it has been checked.

Frontier The most likely outcome is neither replacement nor defeat but absorption. Heterogeneous-agent optimising models already took the distributional content; production-network economics already took the propagation mechanism. A plausible end state is that the agent-based programme supplies components to a mainstream that does not adopt its philosophy — a scientific success and a rhetorical defeat, which the field’s own literature would record as the latter.

These horizons track the evidence base for the modelling paradigm, not the economies being modelled.

• 10 yr: Frontier Either a scored, real-time comparison exists for at least one country — most plausibly Canada, where both entrants have institutional owners in the same building — or the literature is still citing the Austrian and 38-country results, which will by then rest on windows ending in 2019 and 2013. Simulation-based inference keeps improving and the honestly estimable parameter count rises above the current nine-parameter demonstrations; whether it reaches forecasting scale is open. Financial-stability models stay in policy use and keep being cited as evidence for a forecasting claim they do not support.
• 25 yr: Speculative The likely resolution is absorption rather than victory: heterogeneity, production networks and behavioural rules enter mainstream projection models while optimisation and equilibrium are retained. The alternative branch is that a maintained synthetic population becomes a standard statistical-agency product in several countries, at which point entry costs collapse and the comparison gets run many times rather than once. These futures are distinguishable now by whether any statistical agency publishes a synthetic population under a research licence.
• 50 yr: Speculative If distributional and sectoral forecasts acquire their own scored benchmarks, the criterion by which macroeconomic models are judged shifts away from aggregate root mean squared error, and the comparison this brief calls decisive becomes a curiosity answered on the wrong variables. Questions posed across several models in Post Scarcity Economics and Future Labour Markets would then be posed inside one.
• 100 / 250+ yr: Handwave Any claim about which representation of an economy will be in use at that distance is an assertion. What can be said is that the current dispute is about a tractability constraint — how many heterogeneous agents can be estimated rather than merely simulated — and such constraints have historically moved, which makes the specific form of today’s argument unlikely to survive even if its underlying question does.

• Depends on no result another brief on this map has to produce first. What it consumes is the scoring discipline Forecasting Replication and Technology Forecasting develop for other claims, the dated-forecast custody machinery in Decision Markets in Governance, and the observation shared with Economic Resilience that the cheapest useful experiment in an economic-policy field is usually a pre-registration nobody has written.
• Requires (not on this map) six things, none of them a discovery. An independent custodian that accepts a hashed model and a dated forecast and releases neither until the outturn, because without one every comparison is designed by a party that already knows the answer. A pre-analysis-plan standard for model competitions, fixing variables, horizons, loss functions and revision treatment before the first origin, which economics has for randomised trials and not for models. An identification result for behavioural rules in simulated macroeconomic models, settling whether the mapping from parameters to calibration targets separates materially different rule sets. A maintained synthetic population under a research licence, absent which replication and entry stop outside the four or five countries whose statistical systems already support it. Standing compute for repeated calibration of million-agent models, the cost that forced the current calibration down to five moments. And a buyer who pays for a scored forecast rather than a projection narrative.
• Enables sector-level and household-level policy incidence computed from the same run as the aggregate forecast; distributional evaluation of transfers and tax changes without a separate satellite model; propagation analysis for shocks whose sectoral incidence is known in advance, which is what the pandemic work demonstrated; and, if the identification question is answered rather than assumed, policy counterfactuals that survive the objection Lucas raised in 1976.
• Adjacent to Economic Resilience for the policy field that needs this model class and currently reasons without it, Future Capital Markets for the financial-stability branch already in operational use, Multi-Agent Intelligence Systems for the shared simulation machinery, Digital Economies for the high-frequency behavioural data that would let rules be estimated rather than assumed, and Infrastructure Resilience for the network-propagation methods this shares.

Frontier “A fully data-driven agent-based model of Austria beat the official forecasting institutes.” This brief was commissioned against that claim and it does not hold as stated. The published abstract says the model can compete with benchmark VAR and DSGE models out of sample; the working paper compares against VAR, AR and DSGE and contains no comparison with the Oesterreichische Nationalbank, WIFO or the Institute for Advanced Studies. The result is real, and it is a comparison against statistical and structural benchmarks rather than institutional forecasts. The related error is authorship: J. Doyne Farmer is senior author of the later 38-country extension, not of this paper.

Established “DSGE models failed at forecasting, which is why alternatives are needed.” The measured record does not support the premise. Edge and Gürkaynak found the Smets and Wouters model had the lowest root mean squared error for both inflation and growth at longer horizons against Bayesian VAR, Greenbook, Blue Chip and random-walk benchmarks. The honest version of the criticism is the one those authors make themselves: the DSGE model in an absolute sense did a very poor job of forecasting, all three methods are poor, and combining them does not help much. The case for an alternative rests on the absolute standard, not the relative ranking.

Established “Agent-based models can generate any output you like, so they explain nothing.” This is the strongest objection and it is partly correct, which is why the Austrian paper states it against itself: with over-parameterisation and the corresponding degrees of freedom, almost any simulation output can be generated. What has changed is that Bayesian and simulation-based estimation now impose real discipline. What has not changed is that nobody has shown the estimation problem is identified at forecasting scale, so the objection survives in weakened form and 7 names the numerical test that would retire it.

Frontier “Agent-based models escape the Lucas critique.” The argument is that rules estimated from microdata are invariant to policy where aggregate relations are not. The counter-argument is that a rule fitted within one regime encodes that regime’s expectations too, so a fixed-rule simulation misforecasts a regime change for exactly the reason Lucas gave. Neither has been tested, because no agent-based macro model has been asked to forecast across a documented regime change it was not calibrated over.

Established “Central banks have adopted agent-based models.” True and routinely overread. The Bank of England’s 2025 survey covers over twenty central banks, and the great majority of the work is financial stability: interbank contagion networks and housing-market models, with results in financial stability reports since 2016 at the Magyar Nemzeti Bank. One central bank has given the approach a formal place in its official macroeconomic modelling strategy. A count of institutions is not a count of macroeconomic forecasting uses.

Frontier “The 38-country result settles it.” It is the strongest quantitative evidence available and it does not. The comparison is against an AR(1) and against the earlier agent-based model, with no DSGE or Bayesian VAR entrant; the margin over its own predecessor is 65 per cent of Bayes factors; the calibration window ends in 2013Q1; and the evaluation is neither on real-time vintages nor pre-registered. Each is a normal property of a research paper, and jointly they are why a tournament is a different object.

Frontier “More agents means more realism means better forecasts.” The chain has two untested links. No published study reports the forecast-accuracy curve against agent count, and the model with the best externally checkable forecasting record — the United Kingdom pandemic work — was a dynamic disequilibrium input-output model rather than a one-to-one agent population. The ablation in 7 is the direct test and has never been run.

Handwave “Agent-based models predicted the 2008 crisis.” They did not, and the founding documents of the modern programme do not claim they did. Farmer and Foley argued in Nature in 2009 — after the crisis — that the economy needs agent-based modelling, which is a proposal rather than a track record. The claim circulates because the critique of the incumbent and the promotion of the alternative arrived in the same sentence.

Speculative “The pandemic proved the approach.” It proved something narrower and worth having. A lockdown is a shock whose size and sectoral incidence are announced in the policy order, the case where a structural propagation model should beat anything estimated on a history containing no such event. The New York model matching industry-level employment change at a Pearson correlation of 0.82 is a real result about propagating a known impulse, not evidence about forecasting an ordinary quarter.

Frontier “The two paradigms are rivals.” Possibly not. Forecast combination beat both DSGE and Bayesian VAR at all horizons in the Riksbank comparison, and no published study reports the correlation between agent-based and DSGE forecast errors on the same economy. If that correlation is low, the right conclusion from a tournament is an ensemble rather than a winner, and the adversarial framing will have cost the field a decade.

Frontier “The honest position is that one paradigm is clearly ahead.” The honest position is that they have never been compared under conditions that could settle it. One has a fifteen-year real-time forecasting record that is mediocre in absolute terms and well documented; the other has one published head-to-head on one country over a window with no recession in it, a 38-country result scored against weak benchmarks, and a genuine success on a shock whose size was known in advance. That is not a tie declared out of politeness. It is what the record supports, and the experiment that would replace it costs a protocol, a custodian and five years.