The French administration combines devolution and decentralisation within the framework of a unitary state. The State has gradually transferred powers to local authorities. Whilst this decentralisation brings certain decisions closer to citizens, it makes the issue of financial resource allocation particularly sensitive. This is the ‘game of compromise’ analysed by Patrick Le Lidec in the relations between the State and local authorities1.
When budgetary trade-offs reduce central government funding for a particular public policy, one might hope that local authorities will make up the shortfall. Some may well do so. Others will postpone their investments, with consequences for all stakeholders dependent on this funding, starting with the businesses that were expecting orders. Whilst the amount cut is easy to calculate, the actual consequences, both direct and indirect, are less so. Yet it is the sum total of these consequences that determines the true cost of the measure.
To prepare the French government’s 2027 budget, we propose subjecting measures to a numerical simulation capable of calculating their economic effects, as well as modelling behavioural responses and their consequences over time. Starting with a few clearly defined measures would allow us to verify the data, rules and results before broadening the scope. Simulating the entire budget in a single step would multiply the number of interactions and assumptions that are difficult to validate.
Our proposed model goes beyond a mere accounting projection. The aim is to construct a model in which households, businesses and public authorities, represented by computer agents, react according to their expectations, resources and constraints. The decisions of each of these actors alter the situation of the others. A local authority that defers an investment reduces its immediate expenditure, but also alters its suppliers’ order books. The model must make it possible to track this chain of events and then compare several budgetary options without having to wait for their effects to materialise.
Economic simulation already has robust tools at its disposal. EUROMOD, Ines and TAXIPP assess the effects of taxes and benefits on household incomes. The static core of EUROMOD and the Ines model, on their own, do not replicate the adjustments triggered by a reform. However, this limitation does not apply to microsimulation as a whole: EUROMOD can be coupled with behavioural models, whilst TAXIPP formalises, in particular, assumptions regarding the take-up of benefits. France Budget, for its part, offers macroeconomic projections based on aggregate relationships2. The proposed model would complement these approaches by explicitly representing the interdependent decisions of local authorities and their suppliers, and their consequences over time.
Members of Parliament already have access to the annexes to the draft Finance Bill, preliminary assessments and annual performance plans, as well as committee reports, questionnaires and hearings. The National Audit Office also assists them in monitoring the implementation of the budget3. The proposed addition would enable them to examine how a projected saving changes as stakeholders adapt their decisions, and what assumptions account for these discrepancies.
Multi-agent models make it possible to study these interactions between different actors. As early as 2016, the Bank of England outlined their principles and applications to the housing and corporate bond markets. A working paper published in February 2025 notes their increasing use within central banks, both for research and for the analysis of public policy. These models are presented therein as complementary tools, particularly useful for representing the heterogeneity of agents and their interactions4. There is therefore a body of literature and established practices on which to draw when constructing a model suited to French budgetary measures.
Turning disagreement into a useful outcome
To illustrate our approach to studying the consequences of a reduction in public support for local investment, let us imagine a model representing a group of local authorities and the businesses that depend on their contracts. The former would have revenue, cash flow, contractual commitments and infrastructure to maintain. The latter would differ in terms of their size, their dependence on public procurement and their ability to find other clients. One month of simulation would correspond to one month of activity. This scale would make it possible to track payments and deferrals, without attempting to replicate the entire economy.
The simulated population would be constructed using open administrative and statistical data. A synthetic population can preserve distributions and certain observed relationships without faithfully replicating the situation of each real-world actor. However, it is still essential to preserve the relationships that matter. If the model randomly distributes vulnerable suppliers across regions, whilst in reality they are concentrated in a few employment areas, it risks underestimating local difficulties. Simply reproducing a national average is therefore not enough: the model must also preserve the location of the actors and the dependency links that transmit the effects of a reduction in funding.
This raises the question of behaviour: what do agents do when their resources diminish? A local authority may postpone a construction project, cut another item of expenditure or draw on its cash reserves. A company may seek private contracts, reduce its purchases or defer recruitment. These rules must be estimated on the basis of observed behaviour where the data permit. In other cases, several hypotheses must remain in contention. Assuming, for example, that all local authorities postpone their investments would lead to different results from a scenario where some use their cash reserves. This assumption must remain explicit, rather than being buried in the code and treated as a fact.
To avoid these pitfalls, we would distinguish between directly measured parameters, those that are estimated, and those that are explored for lack of a better alternative. The model should respect accounting identities: a payment is an outflow for one party and an inflow for the other. We should also specify what lies outside the scope, in particular private markets or credit conditions. Without this precaution, a fall in revenue for a supplier could be omitted from the calculation or, conversely, counted multiple times as interactions occur.
Before introducing the budgetary reform, we must verify that the computer programme running the simulation works as intended, and then examine how accurately it replicates past data. We would adjust its parameters over one period and, where data permit, set aside another period for validation. In practical terms, this would involve checking whether the model reproduces the payment delays, fluctuations in investment or cash flow difficulties observed. It should also be compared with a simpler method, such as a projection assuming behaviour remains unchanged. If adding a mechanism does not alter the conclusions or improve the model’s ability to replicate the observations, its usefulness would need to be reassessed. Simplification would then involve removing superfluous mechanisms and reducing the number of parameters, whilst retaining those that explain the effects under study.
The added value of the model: testing a reform in advance
Testing a reform begins with a baseline scenario, constructed without the measure under study. Several variants are compared against this baseline, under the same external conditions. This is essential because a simulated deterioration cannot be entirely attributed to the reform if it also appears in the scenario without the reform. The results remain contingent on the decision-making and interaction rules adopted for the agents. Simulation helps to examine assumed causal mechanisms, but it does not, on its own, establish that they accurately describe reality. It is up to analysts to define a validation method and to test these mechanisms against observations.
We would then vary the uncertain parameters and, if the model includes random events, repeat the calculations. Sensitivity analysis would reveal which assumptions actually influence the result. If the net economic impact depends almost entirely on a reaction from local authorities that no one can estimate, the decision-maker must recognise this. It would be misleading to announce a probability of success based on a simple catalogue of scenarios. Simulation frequencies only become relevant probabilities under justified assumptions regarding parameters and uncertainties.
We should also identify the conditions under which the measure ceases to produce the expected effect. A rise in the cost of credit or in payment terms may alter firms’ responses. This ‘stress testing’ approach is already used in finance. Bank stress tests, for example, examine the effects of adverse scenarios on institutions’ ability to absorb losses5. Applying it to a budgetary measure would require its own models. The benefit would be to identify critical dependencies before being forced to finance them as a matter of urgency.
The useful result would therefore go beyond a simple sum of savings. It would specify when these savings materialise, who bears the transition costs, and which stakeholders become vulnerable. Budgetary costs would be distinguished from private losses and effects on economic activity, to avoid misleading comparisons. We would also examine cases where the conclusion changes under a different behavioural rule. A result that holds up under several plausible scenarios deserves more attention than a highly precise estimate obtained using just one.
From theory to practice: how to implement the model
For 2027, a few clearly defined measures would suffice to set this work in motion. Each model should be provided with its assumptions, tests and the versions of the data used. The code and results would be made available to an independent review team, subject to the necessary confidentiality restrictions. This review could be entrusted to a scientific committee set up for this purpose, comprising economists and simulation specialists, and mandated by the Finance Committees of the National Assembly and the Senate. It should be possible to rerun the calculations and understand why two versions differ. Public procurement would benefit from requiring this reproducibility from the outset of the contract, rather than a final presentation containing figures that cannot be verified.
In the United States, the Congressional Budget Office already uses models to inform Congress’s budgetary decisions. In particular, it provides a tool that allows economic assumptions to be modified and their consequences on revenue, expenditure, the deficit and debt to be examined6. This example demonstrates the value of explicit and comparable scenarios; it does not, in itself, constitute a precedent for the multi-agent model proposed here. Parliamentarians would thus have an additional tool with which to scrutinise a projected economic scenario. If it depends on the deferral of local investment, they could examine the simulated effects on the businesses concerned. If the result changes as soon as a poorly understood parameter is altered, they would know what information to request before voting. The minister would remain free to defend his measure. He would simply have to account for his assumptions, rather than presenting the expected figure as a foregone conclusion.
The Opinion is available below
- Le Lidec, P. (2007), ‘The game of compromise: the State and local authorities in the context of decentralisation in France’, Revue française d’administration publique, No. 121–122, pp. 111–130. doi:10.3917/rfap.121.0111. ↩︎
- European Commission, JRC, ‘What is EUROMOD?’; INSEE, ‘The INES microsimulation model’; Institute for Public Policy, TAXIPP documentation, ‘Social and tax system and simulation assumptions’; France Budget, simulator and methodology. ↩︎
- National Assembly, dossier on the draft Finance Bill for 2027 and annexes; factsheet no. 60 on the role of standing committees in scrutinising the Government. ↩︎
- Bank of England, A. Turrell, ‘Agent-based models: understanding the economy from the bottom up’, 2016; A. Borsos et al., ‘Agent-based modelling at central banks: recent developments and new challenges’, Staff Working Paper No. 1122, 2025. ↩︎
- Bank of England, ‘Stress testing the UK banking system: Guidance on the 2025 stress test for participants’, 2025. ↩︎
- Congressional Budget Office, ‘Easy-to-Use Modelling Tools’, in particular ‘How Changes in Economic Conditions Might Affect the Federal Budget, 2026 to 2036: An Interactive Tool’, 2026. ↩︎