Economic Model-Informed Neural Networks (EMINNs)

• Economic Model-Informed Neural Networks (EMINNs) are neural architectures that integrate economic theory and constraints with deep learning, ensuring outputs align with equilibrium conditions.
• By embedding economic primitives and equilibrium models into training layers, EMINNs offer improved interpretability and robust predictive performance in both empirical and theoretical settings.
• EMINNs employ hard and soft constraint methodologies, such as market-clearing layers and regularizations, to enforce economic consistency across micro, macro, and financial applications.

Economic Model-Informed Neural Networks (EMINNs) are a class of machine learning architectures that integrate economic theory and constraints into the design, training, and interpretation of neural networks. Unlike conventional machine learning approaches that prioritize predictive accuracy without imposing structural or equilibrium properties, EMINNs are explicitly constructed to faithfully encode economic models’ primitives, equilibrium conditions, and theoretical restrictions, ensuring outputs are interpretable, robust, and consistent with economic reasoning. This synthesis allows EMINNs to address complex empirical and theoretical questions in econometrics, microeconomics, macroeconomics, and finance by leveraging the representational power of deep learning while preserving essential economic structure (Farrell et al., 2020, Liang, 26 Aug 2025, Gu et al., 2024, Azinovic et al., 2023, Antulov-Fantulin, 5 Dec 2025, Wu et al., 2024).

1. Core Principles and Foundational Architecture

At the highest level, an EMINN consists of a modular architecture that decomposes into two main components (Liang, 26 Aug 2025):

• Primitive Network: Maps observed economic data (e.g., covariates, actions, observables) to model primitives (utilities, tastes, technology coefficients, expectation vectors).
• Theory Layer: Implements the structural economic model, encoding constraints such as equilibrium, utility maximization, market clearing, or policy Euler equations. This layer is either strictly hard-wired in the architecture or softly enforced via regularization terms in the loss.

For example, in the canonical EMINN for heterogeneous treatment response (Farrell et al., 2020):

1. Baseline Model: Data $(Y, T, X)$ with outcome $Y$, policy/treatment $T$, features $X$. Standard parametric M-estimation is

$$\hat\theta = \arg\min_\theta \frac{1}{n} \sum_{i=1}^n \ell(Y_i, T_i; \theta)$$

2. Flexible Parameterization: Replace parameter $\theta$ with a function $\theta(X)$, estimated by a deep neural network. The economic loss is evaluated as

$$\hat\theta(\cdot) = \arg\min_{\theta(\cdot)\in\mathcal{F}_\text{DNN}} \frac{1}{n} \sum_{i=1}^n \ell(Y_i, T_i; \theta(X_i))$$

This pattern recurs throughout EMINNs: one network maps features to primitive economic quantities, which then serve as inputs to analytic layers or loss functions encoding equilibrium or structural constraints.

2. Enforcement of Economic Constraints in Neural Networks

A fundamental innovation of EMINNs is the rigorous imposition of economic constraints during neural network training and architecture specification (Azinovic et al., 2023, Liang, 26 Aug 2025, Antulov-Fantulin, 5 Dec 2025):

• Hard constraints via architectural layers: For instance, market-clearing layers are designed so that, for asset positions $b^i(x)$,

$$\sum_{i=1}^H \mu_i b^i(x) = B, \quad b^i(x) \ge \underline b^i(x)$$

is automatically enforced by parameterizing $b^i$ as a projection onto the feasible set. Box constraints and linear restrictions can be encoded as custom layers or implicit quadratic programming (QP) mappings.

• Soft constraint regularization: Deviations from properties such as monotonicity, concavity, or equilibrium can be penalized in the loss function by terms such as

$$\mathcal{R}_\text{mono}(\phi) = \sum_i \text{ReLU}(u_\phi(z_i) - u_\phi(z_{i+1}))^2$$

for monotonicity of utility, or

$$\mathcal{C}_\text{eq}(\phi) = \|D_\phi(P^*, X) - S_\phi(P^*, X)\|^2 + \alpha\,\text{Var(roots)}$$

for approximate market-clearing.

• Homotopy and staged training: For numerically unstable problems (e.g., multi-asset equilibria), EMINNs use a homotopy scheme, gradually introducing assets, supplies, or constraints, and slowly ramping up penalization or unmasking new policies to stabilize convergence (Azinovic et al., 2023).
• Graph and locality constraints: In agent-based macro or network-finance models, EMINNs structure message-passing or GNN layers to respect only local interactions, conservation laws, and bounded rationality (Antulov-Fantulin, 5 Dec 2025, Cao et al., 2023).

The table below summarizes common economic constraints and their EMINN implementations:

Economic Constraint	EMINN Implementation	Reference
Market clearing	Architectural layer (projection)	(Azinovic et al., 2023)
Budget/conservation law	Hard split in network outputs	(Antulov-Fantulin, 5 Dec 2025)
Monotonicity, concavity	Loss penalties/activation design	(Liang, 26 Aug 2025)
Nonnegativity, zero-at-zero flows	Activation (ReLU), loss penalty	(Antulov-Fantulin, 5 Dec 2025)
Bounded rationality/cognitive lag	Step-size in Neural ODE, memory	(Antulov-Fantulin, 5 Dec 2025)

By construction, outputs of EMINNs, either "hard" or "soft," comply with the mathematical structure of the target economic model.

3. Estimation, Training, and Inference Methodology

EMINNs are trained by minimizing composite losses that combine standard statistical or empirical fit with latent violation metrics for economic structure (Farrell et al., 2020, Yanhao et al., 7 Feb 2025, Wu et al., 2024):

• Main empirical loss: Typically log-likelihood, GMM criterion, or mean squared error:

$\ell(\hat y, y)$

• Economic and equilibrium penalties: Theory regularizers and constraint penalties:

$$\mathcal{R}(\phi, \theta) + \mathcal{C}(\phi, \theta)$$

• Total loss: Aggregate sum with user-determined weights:

$$\min_{\phi, \theta} \frac{1}{n} \sum_{i=1}^n \ell(m_\theta(h_\phi(x_i)), y_i) + \lambda \mathcal R(\phi, \theta) + \gamma \mathcal C(\phi, \theta)$$

• Optimization: Trained by Adam or variant SGD methods. Regularization via dropout, weight decay, early stopping is common.
• Automatic differentiation: Required for both back-propagation and for calculating PDE residuals in models with dynamic structure (e.g., HJB equations) (Wu et al., 2024, Gu et al., 2024).
• Double machine learning and influence functions: For inference on parameters and functionals, EMINNs employ Neyman-orthogonal scores and cross-fitting for robust confidence intervals, e.g.: $$\varphi(Y,T,X) = H(X, \theta_0(X); t^*) - H_\theta(X, \theta_0(X); t^*)\, \Lambda(X)^{-1} \ell_\theta(Y,T; \theta_0(X))$$ With plug-in and cross-validated estimation, root-n normality of target functionals can be established under polynomial rates (Farrell et al., 2020).
• Simulation-trained neural estimators: For models with intractable likelihoods or heavy simulation cost, EMINNs (a.k.a. Neural-Net Estimators) can be trained on simulated moment–parameter pairs, learning to recognize parameter vectors directly from data summaries, achieving accuracy and computational efficiency relative to classical SMM or SMLE (Yanhao et al., 7 Feb 2025).

EMINNs can be applied in a wide variety of empirical workflows: classical supervised learning, moment-matching, PDE residual minimization for continuous-time equilibrium, and graph-based or agent-based ODE rollout for macro/ABM settings.

4. Applications Across Microeconomics, Macroeconomics, and Finance

EMINNs are widely deployed across diverse economic domains, offering substantial advantages in flexibility, interpretability, and scalability:

• Heterogeneous Treatment Effects and Demand Estimation: Flexible DNN parameterizations of individual-level coefficients enable rich non-linear heterogeneity, accurate counterfactual and personalized pricing analysis, and valid statistical inference via influence functions (Farrell et al., 2020). For instance, in consumer choice models, EMINNs uncover marked heterogeneity in price sensitivity and enable policy-relevant optimal pricing.
• Structural Model Parameter Inference: EMINNs trained on synthetic moments achieve lower RMSE and bias than smoothed maximum likelihood estimators, retain robustness to redundant moments, and output accurate uncertainty quantification, as validated in consumer search models and linear time series settings (Yanhao et al., 7 Feb 2025).
• Dynamic Macro and Agent-Based Models: Modeling macroeconomic time evolution with EMINNs allows encoding of locality (GNNs for agent or sector networks), preservation of budget and conservation laws, and support for policy counterfactuals via differentiable ODE rollouts. For example, in country-level GDP forecasting, EMINNs replicate both in-sample and out-of-sample (post-pandemic) behavior, and permit gradient-based evaluation of policy interventions such as trade link disruptions (Antulov-Fantulin, 5 Dec 2025).
• Modeling Financial Networks and Price-Mediated Contagion: Dual-net EMINNs model unobserved liquidations and asset prices, imposing economic monotonicity via parameter clamping and ReLU, recovering hidden liquidation flows and aligning with observed equilibrium prices, without requiring explicit analytic form of inverse demand (Cao et al., 2023).
• Continuous-Time Equilibrium and PDE Solution: Deep-MacroFin and related EMINN frameworks solve high-dimensional PDEs such as HJB, master equations, and coupled algebraic constraints. They support global, non-perturbative solutions to dynamic macro models with many heterogeneous agents, circumventing curse-of-dimensionality present in grid-based methods (Gu et al., 2024, Wu et al., 2024). EMINNs yield accurate recovery of equilibrium value and policy functions in Krusell–Smith, Khan–Thomas, and macro-finance models (MAPE ≈ 2–3%, training wall-clock times 2–6 hours for high-dimensional state spaces).
• Homotopy and Multi-Asset Models: EMINNs address multi-asset equilibria instability by staged homotopy—incrementally introducing additional margins (assets, constraints), ensuring convergence and interpretability for overlapping generations models with credit and portfolio choice (Azinovic et al., 2023).
• Behavioral and Learning Models: In environments with learning agents (as opposed to rational expectations), EMINNs model agent policies via shallow networks trained on history and experience, efficiently capturing learning traps, departures from canonical benchmarks (higher MPC, greater hand-to-mouth), and resulting wealth/consumption distributions (Kuriksha, 2021).

5. Interpretability, Causal Analysis, and Extensions

EMINNs are constructed to yield interpretable primitives and support structural policy analysis:

• Interpretable latent representations: Hidden layers often correspond to economic primitives (utility, marginal propensity, liquidation, etc.), supporting semantic inspection and diagnostics (Liang, 26 Aug 2025, Cao et al., 2023).
• Causal inference and counterfactuals: By encoding mechanistic constraints, EMINNs allow gradient-based analysis of counterfactuals, policy interventions, and marginal comparisons, e.g., $\partial \hat X / \partial \tau$ for tax or trade shocks (Antulov-Fantulin, 5 Dec 2025).
• Adversarial search and theory discovery: EMINNs can be coupled with adversarial generators to probe the limits and anomalies of economic theory—training generators to design games or scenarios that “break” the baseline model, exposing systematic failures or prompting new theory refinement (Liang, 26 Aug 2025).
• LLM integration: EMINNs can leverage LLMs to generate hypotheses, label latent clusters, or act as synthetic experimental subjects for economic models (Liang, 26 Aug 2025).
• Scalability and generalization: Architectural modularity, compositional loss, and hybrid hard/soft constraint encoding enable adaptation to arbitrary economic models, PDE systems, and agent architectures, including up to 50-dimensional PDEs in macro-finance settings with polynomial scaling in computational cost (Wu et al., 2024).

6. Theoretical, Empirical, and Computational Guarantees

EMINN research establishes a range of theoretical and empirical performance results:

• Convergence rates and non-asymptotic bounds: Under regularity and smoothness conditions, EMINNs with ReLU networks achieve

$$\|\widehat\theta_k-\theta_{0k}\|^2_{L_2(X)} = O(n^{-p/(p+d_C)}\log^8 n)$$

for $C^p$ parameter functions (Farrell et al., 2020).

• Influence function–based inference: Valid confidence intervals can be constructed using Neyman-orthogonal scores when network estimation achieves first-stage rate $o(n^{-1/4})$ (Farrell et al., 2020).
• Robustness to redundant covariates/moments: EMINNs demonstrate insensitivity to spurious features, in contrast to classical estimators whose RMSE increases with irrelevant moments (Yanhao et al., 7 Feb 2025).
• PDE and HJB solution accuracy: Residual losses drop 3–6 orders of magnitude in macro-finance PDEs, with function errors below $10^{-5}$ in equilibrium and transition path recovery (Gu et al., 2024, Wu et al., 2024); curse of dimensionality can be mitigated for systems with up to 200 state variables.
• Empirical performance: Out-of-sample fit matches or exceeds classical models, with substantial gains in transfer, interpretability, and causal generalization.

A plausible implication is that EMINNs are uniquely positioned for large-scale, theory-consistent economic analysis on high-dimensional, heterogeneous data, and may provide both predictive and scientific insight not attainable by conventional machine learning or purely analytic modeling.

7. Limitations and Future Directions

Although EMINNs demonstrate significant promise, several frontiers and challenges remain:

• Computational sensitivity: Homotopy, initialization, and training schedules can affect convergence, especially in high-dimensional, multi-equilibrium environments (Azinovic et al., 2023).
• Selection of constraint weights: Choice of regularization strength ($\lambda, \gamma$) crucially impacts performance; further research on principled selection is ongoing (Liang, 26 Aug 2025).
• Interpretability in large, deep models: While primitives are interpretable in low-dimension cases, complex EMINNs may require external tools (e.g., LLMs, feature attribution) for transparency (Liang, 26 Aug 2025).
• Extensions to partially observed systems: Several EMINN architectures are designed for full information; semi-supervised and missing-data settings remain active areas (Cao et al., 2023).
• Integration with symbolic and programmatic model generation: The modularity of EMINNs is compatible with programmatic and symbolic model discovery, representing a nexus of machine learning–enhanced economic science (Liang, 26 Aug 2025).

EMINNs thus represent a significant methodological advance—embedding the rigor of theory and equilibrium into the flexibility of modern machine learning, and enabling both more robust predictive modeling and deeper structural understanding across the economics discipline.