PaperlayerView exact revision on arXiv

Source-linked AI summary

Towards Credible Agent-Based Policy Simulations: Disentangling Opportunities and Preferences in a Financial Inclusion Case Study of Egypt

Alba Aguilera, Georgina Curto, Nardine Osman, Ahmed Al-Awah

arXiv:2610.04515v1cs.AIcs.MA

TL;DR

Agent-based policy simulations need empirical grounding and credibility for their intended use, yet contextual constraints and preferences complicate calibration and interpretation. This paper presents a Capability Approach-aligned framework and applies it to financial inclusion in Egypt, using fixed opportunities and calibrated preferences to distinguish possible sources of group gaps.

  • Problem

    Policy simulations need empirically grounded, transparent assumptions, parameters, and outputs, while financial-inclusion gaps reflect both opportunities and preferences across groups.

    Source: Introduction · pp. 1–2

  • Method

    The framework maps an agent-based model to the Capability Approach, builds its population from data, and calibrates group-specific preferences while holding feasibility constraints fixed.

    Source: Introduction · p. 2, Agent population and Environment · p. 5

  • Results

    Calibration reduces weighted root mean square error from 0.101 ± 0.001 to 0.068 ± 0.003, while the analysis distinguishes opportunity gaps from preference-based interpretations across groups.

    Source: Results and Parameter Plausibility · pp. 12, 15

  • Takeaways & Limitations

    The case study turns calibration into a source of explicit, group-specific hypotheses that can be assessed by domain experts and interpreted alongside gaps between opportunities and outcomes.

    Source: Results and Parameter Plausibility · p. 15, Conclusions and Future Work · p. 17

    The model assumes calibrated preferences remain fixed over the policy horizon unless policies explicitly change them, and represents some SME actions as one-time transitions.

    Source: Limitations and Ethics · p. 17

Abstract

from arXiv · show

Credibility is a central topic for agent-based models intended to support policy-making. Simulations must not only represent the target scenarios and their core dynamics but also demonstrate that their assumptions, parameters, and outputs are empirically grounded and sufficiently accurate for their intended use. This paper addresses this challenge by presenting a general modelling framework, aligned with the Capability Approach, for building credible policy simulations that rely on data and domain-expert knowledge. It then demonstrates how it can be contextualised and implemented to study the social challenge of financial inclusion in Egypt, building on an agent-based model that represents heterogeneous individuals and firms behaving according to their financial states, barriers, opportunities, and preferences. The model is fitted to real-world data in two stages, initialisation and calibration, which respectively build representative synthetic populations and estimate behavioural parameters. By fixing the feasibility parameters, which determine agents' opportunities, and calibrating preference parameters across different population groups, we are able to distinguish and analyse the role of institutional and social barriers in the system, as well as the role of agents' motivations and priorities. This calibration stage provides transparent and group-specific hypotheses about the drivers of observed financial-inclusion gaps, which can further be analysed as gaps between agents' opportunities and realised outcomes, a very relevant insight for policy-making. This paper is thus a step towards improving the credibility and usefulness of policy simulations, strengthening the relationship between the model, the real target system, and the stakeholders who will use it. The code is available at: \url{https://www.comses.net/codebase-release/df8383cb-f49b-4f09-8cce-73603b59adcc/}.

Introduction

The paper presents a Capability Approach-aligned framework for empirically grounded policy simulations and demonstrates it through a financial-inclusion model for Egypt. The framework distinguishes agents’ opportunities from their choices and draws on data and domain-expert knowledge to interpret group-specific gaps.

Introduction

  • Agent-based models can show for whom policies change outcomes and through which mechanisms, but their credibility depends on empirically grounded assumptions, parameters, and outputs.Their stochastic outputs, large parameter spaces, and partly unobserved contextual mechanisms make calibration and validation difficult.Source: Introduction · pp. 1–2
  • The framework uses the Capability Approach to distinguish whether people have the opportunity to act from whether they choose to act.This distinction matters because aggregate outcomes alone may not reveal whether a group lacks opportunity or does not prefer an outcome.Source: Introduction · p. 2
  • The framework has four stages: map the model to the Capability Approach, initialise agents and their environment from data, calibrate preferences, and design validation with domain experts.The proposed validation includes submitting calibrated preferences as hypotheses for expert review.Source: Introduction · p. 2
  • The Egypt case study addresses persistent financial-inclusion gaps across groups and services, including account ownership, digital use, saving, borrowing, and firm formalisation.The model represents individuals and firms with heterogeneous financial states and barriers, and calibrates preferences under fixed opportunity constraints.Source: Introduction · pp. 2–3

Related Work

Related work frames credibility as a challenge of calibration, validation, and stakeholder communication. The paper differentiates its CA-aligned model through stochastic, interpretable rules, real-world data, scalability, and planned expert review.

Related Work

  • Unlike CA-aligned work using reinforcement learning to optimise behaviour, this paper uses stochastic action rules shaped by feasibility and preference.The authors describe the rules as an interpretable baseline that could later be replaced or benchmarked against more structured decision models.Source: Related Work · p. 3
  • The paper extends this CA-aligned line with greater scalability, real-data-based non-binary feasibility and preferences, and planned domain-expert review.The review is intended to address both fixed feasibility assumptions and calibrated preference movements.Source: Related Work · p. 3
  • Prior work treats calibration as difficult because ABMs can be stochastic, computationally costly, high-dimensional, and affected by equifinality.Existing approaches include simulated minimum distance, Bayesian estimation, and likelihood-free methods; this paper uses a constructed binomial likelihood from repeated simulations.Source: Related Work · p. 3
  • ABM validation has no universal standard because models differ in their agents, rules, interactions, and emergent outputs.Related work therefore frames validation as structured evidence about the model, target system, and stakeholders, including mechanism plausibility and reproduction of observed patterns.Source: Related Work · pp. 3–4
  • Policy-oriented credibility also depends on traceability, documentation, stakeholder communication, and a clear statement of intended use.Face validity presents model assumptions and behaviour to knowledgeable people for plausibility assessment within a broader validation process.Source: Related Work · p. 4

Methodology through Implementation

The framework maps the Capability Approach into a staged simulation pipeline, separating agents’ opportunities from preferences and realised outcomes. In the Egypt implementation, heterogeneous individuals and firms interact with a context shaped by institutions, while validation remains a plan for future work.

Methodology through Implementation

  • The framework proceeds from Capability Approach mapping to data-based population initialisation, preference calibration against observed targets, and planned domain-expert review.The review and later validation stages have not been executed in this paper.Source: Methodology through Implementation · p. 4

Agent population and Environment

  • The model represents individuals and SMEs across overlapping policy-relevant groups, enabling analysis of intersectional differences in opportunities and outcomes.Institutions and ecosystem supporters shape feasibility through context parameters; policymakers are prospective users, not simulated agents.Source: Agent population and Environment · pp. 5–6

Feasibility, Preference and Choice

  • Feasibility parameters encode fixed personal and environmental constraints, while calibrated preferences capture how desirable actions are once feasibility is considered.Policy multipliers alter opportunities; transition actions are scaled to accumulate gradually rather than saturate population outcomes immediately.Source: Feasibility, Preference and Choice · pp. 6–8
  • Feasibility is assessed against a threshold of 0.5, and some actions depend on earlier transitions, such as account ownership enabling formal saving or borrowing.The threshold is a modelling choice; feasibility probabilities combine barriers and contextual constraints.Source: Feasibility, Preference and Choice · p. 7, Evaluation: Opportunities and Outcomes · p. 8

Evaluation: Opportunities and Outcomes

  • Agents act monthly under stochastic rules: sampled feasible actions update their states, while infeasible selections leave states unchanged and register as deprived opportunities.Realising an action can change feasibility of later actions, such as account ownership making digital payments or saving feasible.Source: Capability Approach Mapping · p. 5, Evaluation: Opportunities and Outcomes · p. 9
  • Opportunity gaps measure available actions not realised; large gaps indicate unconverted opportunities, while jointly low opportunities and outcomes indicate structural deprivation.The evaluation aggregates actions into financial-inclusion-specific central capabilities and uses equal positive action weights in this prototype.Source: Evaluation: Opportunities and Outcomes · p. 9

Initialisation

Initialisation uses empirical and contextual data to build synthetic populations and represent their financial environment. The resulting prototype contains 20,000 people and 4,318 firms.

Initialisation

  • Initialisation draws on micro- and macro-level datasets and reports from institutionally verified sources.Table 2 lists the data sources used for this stage.Source: Initialisation · p. 9, Table 2 · p. 10
  • Synthetic individuals represent people over 16, with socio-demographic, labour, financial-inclusion, usage, and perceived-barrier information.Populations are sampled using a Bayesian network model and individual-level survey microdata.Source: Initialisation · pp. 9–10
  • Synthetic SMEs capture household non-agricultural enterprises, mainly small businesses, with firm characteristics and financial information.When possible, each enterprise is linked to a synthetic individual manager.Source: Initialisation · p. 10
  • The validated prototype contains 20,000 synthetic people and 4,318 synthetic firms.The populations were compared with source microdata using marginal and joint distributions.Source: Initialisation · p. 10
  • Infrastructure parameters represent physical and digital access, with density factors scaling the feasibility of access-dependent actions.Account-opening feasibility is scaled by 0.50 + 0.50 βI; infrastructure expansion raises feasibility toward baseline.Source: Initialisation · pp. 10–11
  • Regulatory parameters represent fees, documentation strictness, and lending guarantees as qualitative levels mapped to feasibility factors.For account opening, fees and documentation scale feasibility by (1 − 0.30 βf)(1 − 0.25 βd), with fees weighted slightly more.Source: Initialisation · p. 11

Calibration

Calibration holds feasibility fixed while estimating preferences from empirical targets, treating posterior shifts as behavioral hypotheses for expert review. The results distinguish preference-related shortfalls from possible opportunity constraints and identify gaps between opportunities and realized outcomes.

Calibration

  • With feasibility fixed, posterior preference movements are interpreted relative to priors; policy scenarios hold preferences fixed and vary opportunities.The interpretation assumes preferences remain approximately stable over the policy horizon considered.Source: Calibration · p. 11

Bayesian Calibration Procedure

  • The Bayesian procedure defines preference priors, simulates candidate parameters, evaluates target likelihoods, and samples the posterior with Metropolis–Hastings.Parameters are estimated on the log scale to preserve positivity, with data-informed prior centers for complement and subgroup-specific actions.Source: Bayesian Calibration Procedure · pp. 11–12

Targets and Likelihood

Results and Parameter Plausibility

  • Weighted root mean square error fell 32.7%, from 0.101 ± 0.001 under the prior baseline to 0.068 ± 0.003 after calibration; SME targets remained harder to reproduce.Mean absolute residuals were 0.026 for individual targets and 0.077 for SME targets; about half of preference parameters had standard deviations above 0.15.Source: Results and Parameter Plausibility · p. 12
  • For individuals, lower digital-payment use is consistent with lower preference among women, while rural digital-payment and borrowing gaps are consistent with lower opportunities.The rural interpretation points to possible unmodelled barriers including device access, affordability, connectivity, or literacy.Source: Results and Parameter Plausibility · p. 13
  • For SMEs, informality and low digital-sales uptake are consistent with lower general-firm preferences, whereas low licensing among smallest, rural, and women-led firms is consistent with opportunity constraints.The latter interpretation points to possible registration costs, administrative complexity, documentation requirements, distance, or weak enforcement incentives.Source: Results and Parameter Plausibility · pp. 13–14
  • The sharp downward movement in general mobile-money adoption preference is treated as a correction for baseline wallet overproduction, not as a behavioral finding.Because this is a general preference movement rather than a subgroup deviation, the subgroup-versus-general interpretation does not apply.Source: Results and Parameter Plausibility · p. 14
  • Realized outcomes fall below opportunities across capabilities; individual gaps are widest for autonomy and digital literacy, while firms show large formalization and digital-literacy gaps.Youth trail other individual groups on opportunities and outcomes; low-capital firms have fewer employment-related opportunities because hiring is barely feasible in the model.Source: Results and Parameter Plausibility · pp. 15–16, Figure 5 · p. 15

Validation

The paper proposes expert review as the first stage of a broader validation strategy, not as a completed validation. Because feasibility assumptions are fixed during calibration, subgroup explanations remain hypotheses conditional on those assumptions.

Validation

  • Calibration does not validate behavioral explanations: subgroup outcomes may reflect limited opportunities, preferences, or both.The resulting explanations are conditional on the assumed opportunity structure and require external assessment by experts familiar with the modeled context.Source: Validation · p. 16
  • The proposed first validation stage is structured expert review of fixed feasibility assumptions and calibrated preference movements.Experts are asked to accept, revise, or reject explicit calibration hypotheses and identify missing mechanisms.Source: Validation · p. 16
  • The planned panel covers central banks, gender- and SME-finance specialists, and Egypt-based field researchers.Experts receive descriptions of agents, opportunities and realized actions, and the behavioral hypotheses generated by calibration.Source: Validation · p. 16

Limitations and Ethics

The framework’s results depend on numerical and conceptual assumptions, including how preferences and SME data are represented. The authors also caution that underlying data and modeling assumptions may carry biases, so the tool is intended to support probabilistic rather than deterministic policy judgments.

Limitations and Ethics

  • SME construction preserves broad manager–enterprise relationships but does not recover every unobserved correlation, and makes sample size endogenous.Digital sales indicate digital market presence rather than direct evidence of digital-payment acceptance; worker insurance is measured only among employer SMEs.Source: Limitations and Ethics · p. 17
  • Numerical baseline and multiplier values for action feasibility are assumed, as are prior distributions for preference parameters.The paper notes additional assumptions in constructing the synthetic SME population because available firm data differ substantially from individual-level data.Source: Limitations and Ethics · p. 17
  • The model assumes rule-based choices prioritized by feasibility and preference, preferences inferred from outcomes and fixed over the policy horizon.Some SME outcomes are modeled as one-time transitions rather than repeated yearly choices.Source: Limitations and Ethics · p. 17
  • Data and modeling assumptions may reflect biases that affect persistently underrepresented groups, including people with disabilities.The tool is intended for probabilistic prescriptions, not deterministic predictions, and to support rather than replace policymakers.Source: Limitations and Ethics · p. 17

Conclusions and Future Work

The paper presents a Capability Approach-aligned framework and demonstrates it in an Egyptian financial-inclusion case study. The case study uses calibration to generate policy-relevant hypotheses about gaps, while expert assessment and further validation remain future work.

Conclusions and Future Work

  • The framework combines Capability Approach mapping, data-based initialization, calibration against observed outcomes, and a structured validation strategy.These stages aim to strengthen transparency, interpretability, and empirical grounding in policy simulations for real-world decision-making.Source: Conclusions and Future Work · p. 17
  • The Egypt case study distinguishes gaps more likely associated with limited opportunities from those better explained by behavioral preferences.Calibration becomes a hypothesis-generation process, producing explicit behavioral hypotheses for subsequent domain-expert assessment.Source: Conclusions and Future Work · p. 17
  • The paper proposes expert assessment of calibration hypotheses rather than claiming validated behavioral explanations.Future work will implement expert validation and extend the approach through historical input–output validation and benchmarking.Source: Conclusions and Future Work · p. 18

Model and code availability

The authors make the model and materials needed to inspect and reproduce the reported analyses available in a public repository.

Model and code availability

  • The repository includes the model, synthetic-population generator, scripts, an ODD model description, baseline and calibrated configurations, and scripts reproducing every figure and table.The reported materials are available through the cited COMSES codebase release.Source: Model and code availability · p. 18

Appendix A: Numerical Assumptions

The appendix documents the numerical assumptions used for feasibility and preference parameters. Feasibility is fixed before preferences are applied, while preference priors provide starting values for calibration.

Appendix A: Numerical Assumptions

  • Feasibility assumptions define action feasibility before preference parameters are applied, with complement feasibilities set to 1.00 unless policies change them.Final feasibility is clipped to [0, 1], and qualitative context-barrier levels map to 0.3, 0.6, and 0.9.Source: Table 4 · p. 19
  • In the 20,000-person synthetic population, about 63% of individuals report at least one active account-access barrier.For individuals without an active reported barrier, feasibility is informed by the baseline.Source: Table 4 · p. 19
  • Reference no-policy feasibility multiplies the action baseline by applicable barriers, structural conditions, context terms, and precalibration baseline scale, using α = 1 before clipping to [0, 1].Source: Appendix A · p. 19
  • Table 5 reports remaining barrier multipliers for SMEs and context-barrier multipliers.Source: Table 5 · p. 20
  • Preference-prior means θ0 are listed by agent type, action set, and mathematical parameter, with justifications based on observed shrinked ratios in the synthetic population and SME data.Table 6 describes a 20,000-person synthetic population and a 4,318-row synthetic SME; prior values indicate starting tendencies, not final calibrated results.Source: Table 6 · p. 21, Appendix A · p. 21
Loading 2610.04515v1…