Papers
Topics
Authors
Recent
Search
2000 character limit reached

Deep Learning for Solving and Estimating Dynamic Models in Economics and Finance

Published 14 May 2026 in econ.GN | (2605.14493v1)

Abstract: This script offers an implementation-oriented introduction to deep learning methods for solving and estimating high-dimensional dynamic stochastic models in economics and finance. Its starting point is the curse of dimensionality: heterogeneous-agent economies, overlapping-generations models with aggregate risk, continuous-time models with occasionally binding constraints, climate-economy models, and macro-finance environments with many assets and frictions generate state and parameter spaces that strain classical tensor-product grid methods. The exposition is organized around four complementary methodologies. Deep Equilibrium Nets embed discrete-time equilibrium conditions into neural-network loss functions. Physics-Informed Neural Networks approximate continuous-time Hamilton--Jacobi--Bellman, Kolmogorov forward, and related partial differential equations. Deep surrogate models provide fast, differentiable approximations to expensive structural models, while Gaussian processes add a probabilistic layer that quantifies approximation uncertainty; together they support estimation, sensitivity analysis, and constrained policy design. Gaussian-process-based dynamic programming, combined with active learning and dimension reduction, extends value-function iteration to very large continuous state spaces. Applications span representative-agent and international real business cycle models, overlapping-generations and heterogeneous-agent economies, continuous-time macro-finance, structural estimation by simulated method of moments, and climate economics under uncertainty. Companion notebooks in TensorFlow and PyTorch invite hands-on experimentation. These notes are a deliberately subjective and inevitably incomplete snapshot of a rapidly evolving field, aimed at equipping PhD students and researchers to engage with this frontier hands-on.

Authors (1)

Summary

  • The paper introduces novel deep learning architectures like DEQNs and PINNs that embed dynamic model conditions into network loss functions.
  • It demonstrates significant empirical improvements in solving high-dimensional economic models, reducing Euler errors to sub-0.1% levels.
  • It presents GP-based surrogates and active learning strategies, enabling efficient uncertainty quantification and policy design in finance.

Deep Learning for Solving and Estimating Dynamic Models in Economics and Finance

Overview and Motivation

This monograph provides a detailed, practical synthesis of deep learning methods tailored for high-dimensional dynamic stochastic models commonly encountered in economics and finance (2605.14493). The motivating challenge is the curse of dimensionality inherent in state, control, and parameter spaces of modern quantitative models—settings such as heterogeneous-agent economies, overlapping-generations models with aggregate and idiosyncratic risk, dynamic portfolio choice, and climate-macro integrated assessment models cannot be feasibly approximated using classical tensor-product grid methods. The text is directed at advanced graduate students, researchers in central banks and policy institutions, and quantitatively trained applied economists seeking to engage directly with these computational frontiers.

Core Methodological Contributions

The exposition is structured around four complementary deep learning methodologies:

  • Deep Equilibrium Nets (DEQNs): Discrete-time equilibrium conditions (e.g., Euler equations) are embedded directly into neural-network loss functions, replacing classical solution procedures with end-to-end unsupervised minimization via stochastic gradient descent. This approach sidesteps the need for externally generated solution datasets, making the network itself responsible for driving residuals (violations of optimality/environmental conditions) to zero. Applications are given to RA and HA economies, OLG models, and international RBC benchmarks.
  • Physics-Informed Neural Networks (PINNs): For continuous-time models, PINNs directly approximate solutions to complex PDEs—Hamilton–Jacobi–Bellman and Kolmogorov Forward equations—by embedding them as differentiable residuals in the training objective. This enables mesh-free, high-dimensional approximation of policy/value functions without explicit conditional expectation computations. PINNs are compared to classical finite-difference methods and are extended by incorporating features from recent advances such as Economic Model Informed NNs.
  • Deep Surrogate Models with Gaussian Processes (GPs): Surrogates (e.g., fully-connected MLPs, deep kernel learning GPs) are trained to approximate expensive simulation models, enabling fast, differentiable and uncertainty-quantified evaluation for downstream applications including parameter estimation, policy design under constraints, and sensitivity analysis. When combined with GPs, approximate Bayesian inference and global exploration of structural models become computationally tractable.
  • GP-Based Dynamic Programming with Active Learning: Embedding GPR into value function iteration algorithms enables sample-efficient solution of dynamic programming problems, leveraging active subspace methods for dimension reduction and guiding exploration in settings with hundreds of continuous state variables.

A unifying theme is the translation of model-driven economic structure—often unsupervised, i.e., not tied to labeled data—directly into neural loss landscapes or supervised surrogate training pipelines, resulting in broader applicability and better computational scaling for high-dimensional problems.

Numerical Results, Empirical Claims, and Implementation Orientation

Substantial empirical evidence is documented. For example, on high-dimensional IRBC and OLG economies with many shocks or agent types, DEQNs solve models with sparse validation Euler errors on the order of 10−410^{-4} to 10−310^{-3} (interpretable as sub-0.1% consumption mispricing). Surrogate and GP approaches yield rapid, differentiable oracles for parameter estimation in structural SMM pipelines and optimal policy design, demonstrated on both macro/asset pricing and climate IAMs. The text highlights practical recipes—including choice between quadrature and stochastic expectation computation, path- versus residual-based training, and loss normalizations—grounded in strong empirical comparisons and open access working code (TensorFlow/PyTorch).

Contradictory or ambitious claims in the literature are carefully addressed: rather than positing deep learning as a silver bullet, the text emphasizes the importance of economic structure in guiding neural function approximation and the limits of overparameterization in validation. A critical numerical implication is that, for Barron-class functions, deep networks with sufficient capacity avoid the exponential cost of grids, achieving dimension-independent L2L^2 approximation rates, whereas grids and classical polynomials quickly become infeasible.

Theoretical Implications and Open Problems

From a theoretical perspective, the work clarifies that mesh-free, ergodic-distribution-based collocation via neural networks is admissible as long as residuals are reported on independent (out-of-training) state-trajectories and not solely on the economic regions visited during training. The methodology is robust to the double-descent generalization regime commonly demonstrated in deep learning, but emphasizes the necessity of rigorous residual diagnostics and validation-by-simulation rather than parameter counting alone.

Open problems raised include the need for:

  • Off-the-shelf uncertainty quantification pipelines for DEQN/PINN dynamic equilibria, leveraging Bayesian deep learning or high-dimensional GP surrogates.
  • Methods for robustly combining policy/agent-centric and operator-centric function approximation (e.g., hybridizing PINN and DEQN techniques).
  • Infrastructure for scalable, parallel training that integrates with modern AutoML and distributed GPU/TPU environments tailored to the symmetries of economic state spaces.
  • Theoretical results clarifying trade-offs between ergodic-set mesh-free sampling and full-state coverage under non-compact, time-varying, or path-dependent economic environments.

Prospects and Future Directions

Looking forward, the approaches surveyed are particularly well-suited to furthering comparative structural analysis in quantitative macro/finance and structural estimation under model uncertainty or deep nonlinearity. The frameworks are directly extensible to climate-economy applications, robust policy analysis, and multi-agent reinforcement learning models in economics. The integration of probabilistic surrogates, active learning, and high-dimensional dynamic programming remains an active research frontier with potential for substantial impact on both theoretical economics and policy computation.

Conclusion

"Deep Learning for Solving and Estimating Dynamic Models in Economics and Finance" (2605.14493) systematically demonstrates that advances in deep learning architectures, stochastic optimization, and probabilistic surrogate modeling now make feasible the solution and estimation of high-dimensional dynamic economic models previously out of reach for classical numerical methods. The monograph is characterized by technical rigor, explicit methodological comparison, and practical guidance supported by public codebases. It should be regarded as an authoritative resource and a significant stepping stone for the technical development of AI-driven economic modeling and estimation.

Reference:

Paper to Video (Beta)

No one has generated a video about this paper yet.

Sign Up to Generate All Videos Subscribe on YouTube

Whiteboard

No one has generated a whiteboard explanation for this paper yet.

Sign Up to Generate

Explain it Like I'm 14

What this paper is about

This paper shows how modern AI tools—especially deep learning—can help economists and finance researchers solve very complicated “dynamic” models. Dynamic means the models describe how things change over time (like saving and investing every year, or how the climate and the economy affect each other). These models often become too big for old-school methods to handle. The paper explains four AI-based methods that make these big problems manageable, and it gives code so readers can try the methods themselves.

The big questions the paper asks

In simple terms, the paper asks:

  • How can we solve and estimate (fit to data) large, realistic economic and finance models that change over time?
  • How can we get around the “curse of dimensionality,” which is when problems explode in size as you add more factors (like more types of people, risks, or assets)?
  • Can neural networks be taught to respect economic rules (like budget constraints and market clearing) directly while they learn?
  • How can we build fast “stand-ins” for slow models so we can do estimation, test policies, and study uncertainty quickly?

How the authors tackle the problem (with simple analogies)

The paper organizes its approach around four complementary methods. Think of each as a different tool in a toolkit, each best for a certain type of problem:

  • Deep Equilibrium Networks (DEQNs): Imagine teaching a student by giving them rules (like “spend today vs. save for tomorrow”) and grading them on how well they follow the rules. DEQNs train a neural network by directly penalizing it when it breaks core economic equations. Instead of solving the equations first and then training, the network learns by trying to make the rule violations as small as possible.
  • Physics-Informed Neural Networks (PINNs): In continuous-time models (where changes happen at every instant), economics uses special equations (PDEs) that are like the rules of motion. PINNs train a network to produce answers that satisfy these equations. It’s like a coach who doesn’t just show examples but also checks that every move obeys the laws of the game.
  • Deep Surrogate Models + Gaussian Processes (GPs): Some models take forever to run—like a very detailed flight simulator. A “surrogate” is a fast copycat: a neural network that learns to mimic the slow model’s inputs and outputs. Gaussian processes add a “confidence meter,” telling you not just the best guess but also how uncertain it is. This makes it easier and faster to estimate model parameters, run sensitivity tests, and design policies safely.
  • GP-based Dynamic Programming with Active Learning and Dimension Reduction: Dynamic programming is like planning the best choices over time. GP-based methods learn the parts of the problem that matter most (dimension reduction), ask the most informative questions (active learning), and keep track of uncertainty. This can scale to models with hundreds of continuous variables—far beyond what regular grids can handle.

Key idea across all four: don’t separate “economics” from “learning.” Instead, build the economic rules right into how the network learns so it’s trained to follow the rules from the start.

What they find and why it matters

  • It’s possible to solve and estimate much larger, more realistic economic and finance models than before. The new tools sidestep the curse of dimensionality that slows down old grid methods.
  • Embedding the model’s rules into the neural network’s training objective makes learning more accurate and often faster. Rather than learning from examples alone, the network is graded on “how much it breaks the rules,” and it tries to fix that.
  • Fast surrogate models plus Gaussian processes make tasks like parameter estimation, uncertainty analysis, and policy design much quicker. Instead of re-solving a huge model thousands of times, you optimize over a small, fast learner that also tells you how sure it is.
  • GP-based dynamic programming with smart sampling and dimension reduction can handle very large state spaces (up to hundreds of variables). That opens the door to tackling problems that used to be out of reach.
  • The paper emphasizes practical use: it provides diagnostics (ways to check if the solution makes sense) and hands-on code in TensorFlow and PyTorch so students and researchers can learn by doing.

Why it matters: Policymakers, central banks, and researchers can analyze richer models—like ones with many types of households, many assets, or climate risks—more quickly and reliably. That helps them compare policies (e.g., tax rules, climate plans, financial regulations) based on realistic, forward-looking simulations.

Examples of where this helps

  • Macroeconomics: models with many types of people (rich/poor, young/old), multiple shocks, and frictions.
  • Finance: pricing many assets, handling transaction costs, or modeling default risks.
  • Climate–economy models: setting carbon taxes and planning long-term investments under uncertainty.
  • Structural estimation: fitting models to data faster and with a measure of uncertainty.

The potential impact

  • Better decisions, faster: With these tools, researchers can explore more “what if?” scenarios and evaluate policies before they’re implemented.
  • More realistic models: Instead of simplifying away important details (like inequality, rare risks, or many assets), we can include them and still solve the models.
  • More accessible methods: The paper’s code and step-by-step demos make advanced ideas usable for graduate students and practitioners.
  • A moving frontier: The author notes the field is evolving quickly. This work is a practical starting point—expect updates and improvements as the methods and hardware get better.

In short, the paper shows that deep learning can turn previously impossible economic and finance problems into solvable ones, while keeping the economic rules front and center. That means smarter models, faster answers, and better-informed policy choices.

View Paper Prompt  View All Prompts 

Knowledge Gaps

Knowledge gaps, limitations, and open questions

The manuscript presents an implementation-oriented toolbox for DEQNs, PINNs, deep surrogates with GPs, and GP-based dynamic programming, but it deliberately leaves a number of theoretical, algorithmic, and empirical issues unresolved. The following list highlights concrete gaps that future research could address:

  • Convergence theory for DEQNs: establish conditions under which minimizing equilibrium residuals with stochastic gradients converges to true equilibria; characterize approximation, optimization, and sampling error decomposition and provide finite-sample error bounds.
  • Expectation approximation in DEQNs: quantify how Monte Carlo/Quasi-Monte Carlo approximation of conditional expectations affects Euler-equation residuals and gradients; develop variance-reduced, unbiased (or controlled-bias) gradient estimators and guarantees.
  • Multi-residual loss balancing: provide principled, theoretically justified schemes for weighting heterogeneous residuals (Euler, market clearing, transversality, constraints) beyond heuristic methods (e.g., ReLoBRaLo); analyze sensitivity of solutions to weight misspecification.
  • Handling occasionally binding constraints and kinks: design training objectives and smoothing strategies (e.g., Fischer–Burmeister variants) with guarantees for accuracy at non-differentiable boundaries and complementarity conditions; quantify the trade-off between smooth approximations and exact complementarity.
  • Multiple equilibria and selection: develop diagnostics to detect multiple fixed points in residual-based training and methods to control or select equilibria (e.g., continuation, homotopy, regularization, or prior-based selection criteria).
  • Generalization across parameter spaces: study how well DEQNs trained at one calibration extrapolate to nearby structural parameters or policy rules; design curriculum/transfer learning strategies to cover parameter uncertainty or policy search efficiently.
  • Collocation design and adaptive sampling: for both DEQNs and PINNs, derive adaptive, error-indicator-driven strategies for selecting state/collocation points (e.g., residual-based refinement, ergodic-sampling mixtures) and analyze coverage of low-probability but policy-relevant regions.
  • Residual diagnostics and a posteriori error estimates: move beyond pointwise residuals to rigorous a posteriori error bounds for economic quantities of interest (e.g., consumption Euler errors, welfare differentials), including certificates that market clearing/KFE mass conservation holds to specified tolerances.
  • Shape and theory-consistent constraints: integrate and evaluate monotonicity, concavity/convexity, homogeneity, and envelope conditions via architecture (e.g., input-convex networks, lattice models) rather than penalties; quantify gains and potential bias.
  • Stability of dynamic solutions: provide stability tests (e.g., Lyapunov-based, spectral) for learned policy functions and dynamics, especially under long-horizon rollouts and in the presence of approximation noise.
  • PINN convergence for HJB–KFE systems: derive convergence guarantees and sample complexity for joint HJB–KFE training in high dimensions; analyze stability under stiff dynamics and tightly coupled PDEs.
  • Boundary conditions and conservation in KFE: develop architectures and training schemes that enforce non-negativity and exact mass conservation (e.g., via flux-form parameterizations or divergence-free layers) and compare to penalty-based approaches.
  • Choice of PDE loss norms: evaluate alternatives to mean-squared residuals (e.g., weighted residuals, Galerkin/Petrov–Galerkin, Sobolev norms) for PINNs in economic PDEs and quantify their effect on accuracy and stability.
  • Comparison of PINNs vs deep-BSDE vs finite differences: provide systematic benchmarks for continuous-time economic models across methods (accuracy, compute, sample efficiency, robustness to stiffness/constraints) and guidelines for method selection.
  • Operator learning for parametric PDEs: explore DeepONets/FNOs as operators mapping parameters to solutions for HJB–KFE or master equations; quantify generalization across parameter draws and compare with pointwise PINNs.
  • Surrogate-induced estimation bias: formally analyze bias and coverage of structural parameter estimates when replacing the structural model with deep/GP surrogates; develop correction schemes (e.g., SIMEX, bias-aware SMM) and validated uncertainty quantification.
  • Emulator uncertainty in policy design: integrate surrogate predictive uncertainty into constrained policy optimization (e.g., chance constraints, robust optimization) and provide end-to-end guarantees on policy feasibility and welfare under model/estimation error.
  • GP scalability in high dimensions: address O(n3) training cost and kernel design for 100+ dimensional states; evaluate sparse/inducing-point, structured-kernel, and deep kernel learning approaches within GP-based dynamic programming at macro scale.
  • Active subspaces and dimension reduction: develop adaptive, model-aware subspace discovery that updates with policy/parameter changes; quantify approximation error from subspace truncation in dynamic equilibria.
  • Sample-efficient active learning loops: design Bayesian active learning strategies that target regions of high value-function curvature, policy discontinuities, or equilibrium constraints to reduce simulation budgets in surrogate training and GP-VFI.
  • Discrete choices and non-smooth frictions: create architectures/training schemes that handle discrete default, lumpy adjustment, and transaction costs without degrading accuracy near policy discontinuities.
  • Rare events and heavy tails: develop training and importance-sampling techniques that maintain accuracy for tail risks (disasters, climate tails) and quantify their impact on welfare and asset prices.
  • Transition dynamics vs steady state: extend methods and diagnostics from stationary equilibria to transition paths (e.g., after shocks or policy changes), including stability and accuracy guarantees along non-stationary trajectories.
  • Identification and partial identification: establish identifiability conditions for parameters in DEQN/PINN-based estimation and provide tools for partial identification and robust set estimation when moments are weakly informative.
  • Differentiable vs non-differentiable simulators: create hybrid pipelines that combine automatic differentiation with black-box components (e.g., via adjoint methods, surrogate patches, implicit differentiation) and assess resulting estimator properties.
  • Reproducibility and optimizer dependence: quantify sensitivity of learned equilibria to optimizer choice (Adam/AdamW/SGD), random seeds, and mixed precision; develop protocols for robust replication in policy settings.
  • Benchmarks and standardized testbeds: curate open, research-scale benchmark problems (e.g., HANK, OLG-IAM with uncertainty, macro-finance frictions) with agreed accuracy metrics to enable consistent method comparison.
  • Governance for policy deployment: specify validation, stress-testing, and monitoring procedures (model risk, distribution shift) needed before central banks or policy institutions adopt neural solution/estimation pipelines operationally.
View Paper Prompt  View All Prompts 

Practical Applications

Overview

This manuscript introduces four implementation-focused toolkits for high‑dimensional dynamic models in economics and finance: (1) Deep Equilibrium Nets (DEQNs) for discrete-time equilibria, (2) Physics‑Informed Neural Networks (PINNs) for continuous-time HJB/KFE systems, (3) deep surrogates and Gaussian processes (GPs) for fast, differentiable emulation with uncertainty quantification (UQ), and (4) GP‑based dynamic programming (DP) with active learning and dimension reduction. Below are concrete applications derived from these methods, grouped by deployability horizon and linked to sectors, potential tools/workflows, and key dependencies.

Immediate Applications

  • DEQN-based solvers for large DSGE/HANK/OLG models — sectors: public policy, academia, central banking, software
    • What: Replace tensor-product grids with residual-driven neural solvers to compute equilibria in heterogeneous-agent, overlapping-generations, and international real business cycle (IRBC) models that are too high-dimensional for classical methods.
    • Tools/workflows: PyTorch/TF DEQN modules; residual and market‑clearing diagnostics; integration scripts to wrap existing model code; reproducible notebooks for model solution and validation.
    • Assumptions/dependencies: Well-specified economic equations in differentiable form; access to automatic differentiation (AD); validation datasets or moments; moderate GPU/CPU resources; staff familiarity with calibration/diagnostics.
  • PINN solvers for continuous-time HJB/KFE in macro‑finance and option pricing — sectors: finance, insurance, asset management, academia
    • What: Solve HJB and Fokker–Planck equations for continuous-time consumption–savings, portfolio choice with frictions, and derivative pricing without nested conditional expectations.
    • Tools/workflows: PINN HJB/KFE templates; boundary-condition encoding; hybrid FD–PINN cross-validation; automated PDE residual monitors.
    • Assumptions/dependencies: Correct PDE specification and boundary/terminal conditions; numerical stabilization (loss balancing, curriculum, normalization); compute budget for training.
  • Surrogate modeling for structural estimation (e.g., SMM/ABC) — sectors: academia, econometrics vendors, policy institutions
    • What: Train deep/GP surrogates on expensive structural simulators to accelerate estimation, sensitivity analysis, and UQ.
    • Tools/workflows: Oracle–surrogate pipeline; active design of experiments; Deep Kernel Learning for heteroskedastic noise; automated moment‑matching loops with JAX/TF/PyTorch AD.
    • Assumptions/dependencies: Sufficient and representative simulation budgets to cover parameter space; identified moments; governance around emulator error and bias.
  • Surrogate‑based policy search with UQ (e.g., carbon‑tax rules, macro stabilization policies) — sectors: energy/climate policy, public finance, central banks
    • What: Optimize policy rules on trained surrogates of structural models to perform fast, constrained policy design with uncertainty bands.
    • Tools/workflows: Constrained optimizers (e.g., trust-region, projected gradients) over surrogate objectives; Pareto/frontier visualization; scenario libraries; SCC computation pipelines.
    • Assumptions/dependencies: Surrogate fidelity within policy‑relevant domain; clear objective/risk constraints; stakeholder acceptance of emulator‑backed recommendations.
  • Active-learning GP value‑function iteration for high‑dimensional DP — sectors: operations research, energy systems, macro research
    • What: Embed GP regression into VFI with active subspaces to solve dynamic programs with very large continuous state spaces (up to O(102–103) effective dimensions in prototypes).
    • Tools/workflows: GP‑VFI library; acquisition functions for sample efficiency; dimensionality‑reduction diagnostics; checklists for Bellman residuals.
    • Assumptions/dependencies: Smoothness of value/policy functions; stable exploration policies; scalable GP kernel choices or sparse GPs.
  • Fast and differentiable pricing/risk engines — sectors: sell‑side/ buy‑side finance, fintech, risk/treasury
    • What: Train surrogates/PINNs for derivatives, portfolio problems with transaction costs, and risk metrics to enable rapid Greeks, stress testing, and scenario analysis.
    • Tools/workflows: MC/PDE‑generated training sets; autodiff for Greeks; continuous integration tests against benchmark pricers; model‑risk documentation.
    • Assumptions/dependencies: Robust calibration data; model‑risk governance; latency and numerical‑precision requirements of production systems.
  • Sovereign risk and international macro‑finance scenario tools — sectors: IFIs, sovereign debt offices, macro strategy desks
    • What: DEQN‑based solution of multi‑asset, occasionally binding‑constraint environments to explore default probabilities, spreads, and policy scenarios.
    • Tools/workflows: Residual‑based equilibrium solvers; scenario dashboards; sensitivity maps across shocks and parameters.
    • Assumptions/dependencies: Credible structural specification; adequate shock processes; historical calibration data for validation.
  • Portfolio choice with illiquidity and transaction costs — sectors: asset management, robo‑advisory (pilot), private wealth
    • What: PINN/DEQN policies for multi‑asset dynamic allocation under frictions, enabling more realistic rebalancing policies and client‑specific constraints.
    • Tools/workflows: HJB policy heads; constraint‑aware training (KKT/Fischer–Burmeister losses); backtesting harnesses.
    • Assumptions/dependencies: Accurate cost/friction models; sufficient historical/simulated data; compliance and fairness review.
  • Teaching and capacity building with executable notebooks — sectors: academia, central banks, policy schools
    • What: Use the provided notebooks to teach computational macro/finance, accelerate student and staff uptake, and standardize reproducible workflows.
    • Tools/workflows: Lecture–code pairings; sandbox datasets; assignment auto‑graders; GPU‑optional execution.
    • Assumptions/dependencies: Basic Python/ML literacy; institutional support for open‑source tools.
  • Packaged “residual‑driven economics” libraries — sectors: software, consulting, quant platforms
    • What: Productize DEQN/PINN templates with diagnostics, loss balancing (e.g., ReLoBRaLo), and architecture search for fast prototyping.
    • Tools/workflows: API‑first modules for residual definitions; plug‑ins for Dynare/QuantEcon/Econ‑ARK; example galleries.
    • Assumptions/dependencies: Sustained maintenance; version pinning for reproducibility; permissive licensing.
  • Climate economics UQ dashboards — sectors: energy/climate policy, NGOs, ESG analytics
    • What: Serve SCC distributions and policy outcomes from surrogate IAMs; support interactive sensitivity and Pareto‑improvement exploration.
    • Tools/workflows: Pre‑trained IAM surrogates; web UIs with UQ bands; audit trails for parameter draws and priors.
    • Assumptions/dependencies: Transparent parameter choices; communicating epistemic vs aleatory uncertainty; policy‑maker training.
  • Regulatory stress‑testing accelerators — sectors: banking supervision, RegTech
    • What: Link macro surrogates to bank‑level exposure models for rapid, repeated macro‑to‑balance‑sheet scenario sweeps.
    • Tools/workflows: Macro‑scenario generators; mappings to PD/LGD/EAD engines; governance for model chaining.
    • Assumptions/dependencies: Data‑sharing agreements; alignment with supervisory methodologies; explainability and audit requirements.

Long‑Term Applications

  • Real‑time policy “digital twins” of economies — sectors: central banks, finance ministries
    • What: Integrate nowcasting and high‑frequency data with DEQN‑solved HANK/OLG models for near‑real‑time policy analysis and counterfactuals.
    • Tools/workflows: Streaming data ingestion; continual learning for parameters; online UQ dashboards.
    • Assumptions/dependencies: Robust data pipelines; stability under distribution shift; institutional validation and governance.
  • Global IRBC and climate‑macro integration at scale — sectors: international policy, energy/climate
    • What: 100+ country IRBC models coupled to IAM surrogates for coordinated policy analysis (trade, carbon border adjustments, transition risk).
    • Tools/workflows: Modular multi‑region DEQNs; policy‑rule optimizers; scenario orchestration across institutions.
    • Assumptions/dependencies: Cross‑country data harmonization; compute scaling; agreement on structural priors.
  • Certified solvers with a posteriori error bounds — sectors: regulators, risk management
    • What: Embed probabilistic error bars (GP‑UQ) and formal a posteriori bounds into DEQN/PINN outputs for regulator‑grade assurance.
    • Tools/workflows: Conformal prediction; validation suites vs ground‑truth benchmarks; model‑risk reporting templates.
    • Assumptions/dependencies: Advances in theory/practice of NN error certification; regulator acceptance.
  • Micro‑to‑macro integration with high‑dimensional distributions — sectors: public policy, academia
    • What: Combine sequence‑space or histogram‑based HA modules with administrative microdata to improve distributional policy analysis.
    • Tools/workflows: Privacy‑preserving computation (e.g., federated or synthetic data); scalable distribution encoders.
    • Assumptions/dependencies: Data access and privacy constraints; identification of micro‑macro mappings.
  • Operator learning for rapid re‑solves across model families — sectors: software, academia, industry analytics
    • What: Learn solution operators (e.g., via DeepONets/FNOs) so new calibrations or small model changes solve nearly instantly.
    • Tools/workflows: Operator‑level training sets spanning parameter spaces; meta‑learning for fast adaptation.
    • Assumptions/dependencies: Sufficient diversity of training instances; generalization guarantees across regimes.
  • Automated, robust training via NAS and loss balancing — sectors: software, consulting
    • What: Productionize architecture search and adaptive multi‑objective loss balancing (e.g., ReLoBRaLo) for broad model classes.
    • Tools/workflows: Hyperparameter services; failure‑mode detectors; standardized residual scaling.
    • Assumptions/dependencies: Compute budget for NAS; transferability of architectures across problems.
  • Multi‑model ensembles and structural model averaging — sectors: policy evaluation, finance
    • What: Train and combine surrogates for competing structural models to average over model uncertainty in policy design and asset pricing.
    • Tools/workflows: Bayesian model averaging; cross‑model diagnostics; ensemble UQ visualizations.
    • Assumptions/dependencies: Comparable moment sets; principled priors over models; stakeholder buy‑in.
  • Climate‑finance co‑design platforms — sectors: central banks, supervisors, asset owners
    • What: Jointly optimize carbon taxes, transition pathways, and portfolio strategies under deep uncertainty.
    • Tools/workflows: Coupled IAM–risk surrogates; constraint libraries (mandates, budgets); interactive Pareto navigation.
    • Assumptions/dependencies: Shared data standards; multi‑stakeholder governance; credible transition scenarios.
  • Consumer financial planning with dynamic frictions — sectors: fintech, retirement planning (daily life)
    • What: Deploy HJB‑surrogate policies for consumption–savings and retirement decisions with illiquidity and transaction costs to improve advice quality and transparency.
    • Tools/workflows: Mobile/robo‑advice apps with policy explanations and UQ; personalized scenario generators.
    • Assumptions/dependencies: Personal data consent; rigorous backtesting; simple explanations of complex policies.
  • High‑dimensional DP for infrastructure and networks — sectors: energy, mobility, logistics, robotics
    • What: Extend GP‑based DP and active subspaces to optimize grid operations, EV charging, or large‑scale inventory systems with rich state representations.
    • Tools/workflows: Domain‑specific state encoders; safe exploration policies; human‑in‑the‑loop oversight.
    • Assumptions/dependencies: Reliable simulators; safety constraints; problem‑specific kernel/feature engineering.

Notes on feasibility across applications:

  • Methodological fit depends on smoothness/differentiability of equilibrium/PDE residuals; non‑smooth constraints may require specialized losses (e.g., Fischer–Burmeister) or hybrid solvers.
  • Surrogate fidelity hinges on training coverage; active learning mitigates but does not eliminate extrapolation risk.
  • Institutional adoption requires reproducibility, interpretability, and governance (documentation, model‑risk controls, audit trails).
  • Compute budgets matter but classroom‑scale examples run on CPUs; production‑scale deployments benefit from GPUs/TPUs and MLOps practices.
View Paper Prompt  View All Prompts 

Glossary

  • Active learning: A sampling strategy that queries the most informative inputs (often where model uncertainty is high) to improve learning efficiency. "combined with active learning and dimension reduction"
  • Active subspaces: A gradient-based dimensionality-reduction technique identifying important linear directions in high-dimensional inputs. "when combined with active subspaces"
  • Barron-class functions: A function class for which two-layer neural networks achieve dimension-independent approximation rates under bounded Barron norm. "For Barron-class functions, two-layer networks achieve dimension-independent rates"
  • Barron norm: A complexity measure (related to a weighted L1 norm of the Fourier transform) that controls neural approximation rates for Barron-class functions. "bounded Barron norm"
  • Bellman equations: Recursive optimality conditions in dynamic programming that define the value function through a self-consistency relation. "Bellman equations"
  • Bellman operator: The mapping that takes a candidate value function to its one-step dynamic-programming update; its fixed point solves the control problem. "surrogates of the Bellman operator"
  • Curse of dimensionality: The exponential growth of computational cost with the number of state variables, which makes grid-based methods infeasible in high dimensions. "Its starting point is the curse of dimensionality"
  • Deep-BSDE solvers: Deep-learning methods for solving backward stochastic differential equations, often used to tackle high-dimensional PDEs via probabilistic representations. "deep-BSDE solvers"
  • Deep Equilibrium Nets (DEQNs): Neural networks that embed a model’s equilibrium conditions directly into the loss, training by minimizing residuals instead of solving separately. "Deep Equilibrium Nets embed discrete-time equilibrium conditions directly into neural-network loss functions."
  • Deep surrogate models: Learned, fast-to-evaluate approximations of expensive structural models that enable estimation, sensitivity analysis, and policy search. "Deep surrogate models provide fast, differentiable approximations to expensive structural models"
  • Deep uncertainty: Uncertainty about models, parameters, or probabilities themselves, common in long-horizon climate-economy settings. "deep uncertainty"
  • Double descent phenomenon: A non-monotonic test-error curve where error decreases, rises near interpolation, then decreases again as model complexity grows. "double descent phenomenon"
  • Dynamic stochastic general equilibrium (DSGE): A class of macroeconomic models describing intertemporal choices under uncertainty with market clearing and rational expectations. "high-dimensional dynamic stochastic general equilibrium (DSGE) models"
  • Economic Model Informed Neural Networks (EMINNs): Neural networks whose losses include residuals from economic model equations (e.g., master or equilibrium conditions). "Economic Model Informed Neural Networks (EMINNs)"
  • Ergodic sampling: Drawing training points from the long-run (stationary) distribution of states to evaluate residuals or moments representative of the invariant measure. "ergodic sampling"
  • Gaussian processes: Bayesian nonparametric priors over functions that yield predictive means and variances via kernel covariance structures. "Gaussian processes add a probabilistic layer that quantifies approximation uncertainty"
  • Gaussian-process-based dynamic programming: Using Gaussian process regression within value-function iteration, leveraging uncertainty for sample-efficient updates. "Gaussian-process-based dynamic programming, combined with active learning and dimension reduction"
  • Hamilton–Jacobi–Bellman (HJB) equation: The continuous-time PDE characterizing the value function of an optimal control problem. "Hamilton--Jacobi--Bellman"
  • Heterogeneous-agent New Keynesian (HANK) models: Macro models combining nominal rigidities with heterogeneous households to study policy transmission through distributions. "heterogeneous-agent New Keynesian (HANK) models"
  • Integrated assessment models (IAMs): Frameworks that couple climate and economic dynamics to evaluate policies like carbon taxes. "integrated assessment models coupling climate and economic dynamics"
  • Kolmogorov forward equation (KFE): Also called the Fokker–Planck equation; governs the time evolution of probability densities in stochastic systems. "Kolmogorov forward"
  • Market-clearing errors: Residuals measuring supply–demand imbalances that should be zero in equilibrium; used as loss terms in equilibrium training. "market-clearing errors"
  • Pareto-improving: A change making at least one agent better off without making anyone worse off, often used to qualify policy reforms. "Pareto-improving carbon-tax rules"
  • Physics-Informed Neural Networks (PINNs): Neural networks trained to minimize PDE residuals and boundary conditions, enabling mesh-free solution of differential equations. "Physics-Informed Neural Networks approximate continuous-time Hamilton--Jacobi--Bellman, Kolmogorov forward, and related partial differential equations."
  • Simulated Method of Moments (SMM): An estimation approach choosing parameters so model-simulated moments match empirical moments. "structural estimation by simulated method of moments"
  • Tensor-product grid methods: Classical discretization using a dense product grid across state dimensions, suffering from exponential scaling Nd. "tensor-product grid methods"
  • Value-function iteration: An algorithm that iteratively applies the Bellman operator until convergence to the value function. "extends value-function iteration to settings with very large continuous state spaces"
View Paper Prompt  View All Prompts 

Open Problems

We haven't generated a list of open problems mentioned in this paper yet.

Generate Now

Collections

Sign up for free to add this paper to one or more collections.

Sign Up

Tweets

Sign up for free to view the 4 tweets with 77 likes about this paper.

Sign Up for Free