Paper: arXiv 2408.10368

Authors: Yuntao Wu, Jiayuan Guo, Goutham Gopalakrishna, Zissis Poulos

Abstract

In this paper, we present Deep-MacroFin, a comprehensive framework designed to solve partial differential equations, with a particular focus on models in continuous time economics. This framework leverages deep learning methodologies, including Multi-Layer Perceptrons and the newly developed Kolmogorov-Arnold Networks. It is optimized using economic information encapsulated by Hamilton-Jacobi-Bellman (HJB) equations and coupled algebraic equations. The application of neural networks holds the promise of accurately resolving high-dimensional problems with fewer computational demands and limitations compared to other numerical methods. This framework can be readily adapted for systems of partial differential equations in high dimensions. Importantly, it offers a more efficient (5\(\times\) less CUDA memory and 40\(\times\) fewer FLOPs in 100D problems) and user-friendly implementation than existing libraries. We also incorporate a time-stepping scheme to enhance training stability for nonlinear HJB equations, enabling the solution of 50D economic models.

Complexity vs Empirical Score

  • Math Complexity: 9.5/10
  • Empirical Rigor: 2.5/10
  • Quadrant: Lab Rats — high math complexity, limited empirical testing

Why this score: The paper is highly theoretical, focusing on advanced mathematics like high-dimensional PDEs, Hamilton-Jacobi-Bellman equations, and neural network architectures (KANs), with no empirical backtesting or real financial data used; its ‘rigor’ is computational (memory/FLOP benchmarks) rather than financial.

Research Flowchart

  flowchart TD
  A["Research Goal:<br>Solve high-dimensional continuous-time economic models"] --> B["Methodology:<br>Deep-MacroFin Framework"]
  
  B --> C1["Neural Network Architectures:<br>MLPs & Kolmogorov-Arnold Networks"]
  B --> C2["Input Equations:<br>HJB & Coupled Algebraic Equations"]
  
  C1 --> D["Computational Process:<br>Optimization & Time-Stepping Scheme"]
  C2 --> D
  
  D --> E["Key Outcomes:<br>5x less CUDA memory<br>40x fewer FLOPs<br>50D economic model solutions"]

Evidence Card

What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card

Proposes

The paper proposes Deep-MacroFin, a comprehensive deep learning framework for solving high‑dimensional PDEs in continuous‑time economic models.

Source sentence
In this paper, we present Deep-MacroFin, a comprehensive framework designed to solve partial differential equations, with a particular focus on models in continuous time economics.

Assumes 5

  • Time is continuous and the horizon is infinite. model
    Source sentence
    Time is continuous and the horizon is infinite.
  • The model assumes the presence of two state variables: η (experts’ wealth share) and σ (volatility). model
    Source sentence
    Figure 1: System overview. This model assumes the presence of two state variables: η (experts’ wealth share) and σ (volatility), two agents: Ve (experts) and Vh (households), and one endogenous variable: q (price of the capital).
  • Agents can short‑sell when η is below a critical share ηψ. market
    Source sentence
    When η ∈(0, ηψ), agents can short-sell, leading to a nonlinear rise in the capital price q.
  • Loss weights are manually selected for best empirical performance. other
    Source sentence
    we manually select them for best empirical performance, as automatic methods proved less robust.
Show 1 more
  • The neural network approximations satisfy the equilibrium equations without any assumption about the solution itself. model
    Source sentence
    Models that minimize L(θ, T ) ensures the approximated functions satisfy the specified equations and constraints.

Data

  • 100‑dimensional Laplace equation synthetic unknown
    Source sentence
    Specifically, we demonstrate the ability to solve a 100‑dimensional Laplace equation using 4.3GB CUDA memory, 12 GFLOPs, with an average epoch time of 0.16 seconds
  • Lucas orchard benchmark benchmark unknown
    Source sentence
    Figure 5: Lucas orchard benchmark

Costs

Not applicable The paper reports no trading or execution result.

Finds 5

  • Deep‑MacroFin uses 5× less CUDA memory and 40× fewer FLOPs than existing libraries on 100‑dimensional problems. performance
    Source sentence
    fewer FLOPs in 100D problems) and user‑friendly implementation than existing libraries.
  • The framework can solve a 100‑dimensional Laplace equation with 4.3 GB memory and 12 GFLOPs, averaging 0.16 s per epoch. performance
    Source sentence
    Specifically, we demonstrate the ability to solve a 100‑dimensional Laplace equation using 4.3GB CUDA memory, 12 GFLOPs, with an average epoch time of 0.16 seconds
  • A time‑stepping scheme enables convergence of a 50‑dimensional economic model within 10 minutes. performance
    Source sentence
    With time‑stepping and ergodic sampling, the 50‑tree model converges within 10 minutes.
  • KANs provide analytical expressions for critical wealth share ηψ, e.g., ψ = −2.2823 + 3.7254 exp … implying ηψ ≈0.3197. theoretical
    Source sentence
    A key advantage of KAN is its ability to yield an analytical expression for ηψ. For example: ψ = −2.2823 + 3.7254 exp −1.0082(0.6741 −η)2 implies ηψ ≈0.3197 for ψ(ηψ) = 1.
  • Scaling KANs to higher dimensions is challenging due to numerical instability and limited parallelism. negative
    Source sentence
    scaling KANs to higher dimensions remains challenging due to numerical instability with deeper or wider networks and limited parallelism.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 25 related papers). Coverage is partial, so this is neither replication nor refutation.

Code

No repository by the authors found in the paper.

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.

Cited References (50)

  1. Macro, Money and Finance: A Continuous Time Approach (2016) — link
  2. Income and Wealth Distribution in Macroeconomics: A Continuous-Time Approach (2017) — link
  3. Asset Prices and Wealth Inequality (2017) — link
  4. Deep Learning (2016) — link
  5. Workshop Report on Basic Research Needs for Scientific Machine Learning: Core Technologies for Artificial Intelligence (2019) — link
  6. Multilayer feedforward networks are universal approximators (1989) — link
  7. Optimal Control Theory: An Introduction (1970)
  8. Stochastic Controls: Hamiltonian Systems and HJB Equations (1999)
and 42 more
  1. Numerical Treatment of Partial Differential Equations (2007)
  2. Elementary Differential Equations and Boundary Value Problems (2012) — link
  3. Partial Differential Equations and the Finite Element Method (2005)
  4. Numerical Approximation of Partial Differential Equations (2008)
  5. Boundary Element Methods (2020)
  6. Numerical methods for engineers and scientists (2001)
  7. Computational heat transfer (1994)
  8. Deep FDM: Enhanced finite difference methods by deep learning (2023) — link
  9. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations (2019)
  10. Physics Informed Deep Learning (Part I): Data-driven Solutions of Nonlinear Partial Differential Equations (2017) — link
  11. DeepXDE: A Deep Learning Library for Solving Differential Equations (2021) — link
  12. A Deep Learning Based Numerical PDE Method for Option Pricing (2023) — link
  13. Deep Learning for Solving and Estimating Dynamic Macro-Finance Models — → 2305.09783
  14. Macro-financial Modeling in Python: PyMacroFin (2021) — link
  15. A Solution Method for Continuous-Time General Equilibrium Models (2021) — link
  16. ALIENs and Continuous Time Economies (2021) — link
  17. PyTorch 2: Faster Machine Learning Through Dynamic Python Bytecode Transformation and Graph Compilation (2024) — link
  18. Automatic differentiation in PyTorch (2017)
  19. KAN: Kolmogorov-Arnold Networks (2024) — link
  20. A comprehensive and FAIR comparison between MLP and KAN representations for differential equations and operator networks (2024) — link
  21. On the Limited Memory Method for Large Scale Optimization (1989)
  22. Adam: A Method for Stochastic Optimization (2017) — link
  23. Decoupled Weight Decay Regularization (2019) — link
  24. Stochastic differential utility (1992) — link
  25. Substitution, Risk Aversion, and the Temporal Behavior of Consumption and Asset Returns: A Theoretical Framework (1989)
  26. A Local Deep Learning Method for Solving High Order Partial Differential Equations (2021) — link
  27. HoD-Net: High-Order Differentiable Deep Neural Networks and Applications (2022) — link
  28. Neural tangent kernel: convergence and generalization in neural networks (2018)
  29. Optimally weighted loss functions for solving PDEs with Neural Networks (2021) — link
  30. Uncertainty Shocks and Balance Sheet Recessions (2017) — link
  31. A comprehensive study of non-adaptive and residual-based adaptive sampling for physics-informed neural networks (2023) — link
  32. The Lucas Orchard (2013) — link
  33. DGM: A deep learning algorithm for solving partial differential equations (2018) — link
  34. Loss-attentional physics-informed neural networks (2024) — link
  35. Multi-Objective Loss Balancing for Physics-Informed Deep Learning (2021) — link
  36. Solving high-dimensional partial differential equations using deep learning (2018) — link
  37. Solving High-Dimensional Dynamic Programming Problems using Deep Learning (2020) — link
  38. Exploiting Symmetry in High-Dimensional Dynamic Programming (2021) — link
  39. Learning nonlinear operators via DeepONet based on the universal approximation theorem of operators (2021) — link
  40. Ironing, Sweeping, and Multidimensional Screening (1998)
  41. Comment on “Ironing, sweeping, and multidimensional screening’’ (2023) — link
  42. Handling convexity-like constraints in variational problems (2014) — link

1 of 50 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.