Paper: arXiv 2307.06450

Authors: Robert Balkin, Hector D. Ceniceros, Ruimeng Hu

Abstract

In this paper, we propose a numerical methodology for finding the closed-loop Nash equilibrium of stochastic delay differential games through deep learning. These games are prevalent in finance and economics where multi-agent interaction and delayed effects are often desired features in a model, but are introduced at the expense of increased dimensionality of the problem. This increased dimensionality is especially significant as that arising from the number of players is coupled with the potential infinite dimensionality caused by the delay. Our approach involves parameterizing the controls of each player using distinct recurrent neural networks. These recurrent neural network-based controls are then trained using a modified version of Brown’s fictitious play, incorporating deep learning techniques. To evaluate the effectiveness of our methodology, we test it on finance-related problems with known solutions. Furthermore, we also develop new problems and derive their analytical Nash equilibrium solutions, which serve as additional benchmarks for assessing the performance of our proposed deep learning approach.

Complexity vs Empirical Score

  • Math Complexity: 9.5/10
  • Empirical Rigor: 6.0/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper involves dense stochastic calculus, infinite-dimensional HJB equations, and advanced deep learning theory (RNNs, fictitious play), placing it in the high math quadrant. While it validates the method against known solutions and new analytical benchmarks, the lack of real financial data backtesting or specific performance metrics (e.g., Sharpe ratios, transaction costs) limits its empirical rigor compared to purely implementation-focused papers.

Research Flowchart

  flowchart TD
  A["Research Goal<br>Find NE in Stochastic<br>Delay Differential Games"] --> B["Methodology<br>Deep RL via Recurrent Neural Networks"]
  B --> C["Inputs<br>Finance Problems<br>Known & Analytical NE"]
  C --> D["Computation<br>Modified Fictitious Play<br>with Deep Learning"]
  D --> E{"Evaluation"}
  E --> F["Key Finding 1<br>Effective for High-Dim<br>Finance Problems"]
  E --> G["Key Finding 2<br>Validated on<br>Analytical Benchmarks"]

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 a numerical methodology for finding the closed-loop Nash equilibrium of stochastic delay differential games by parameterizing each player’s control with recurrent neural networks and training them via a deep fictitious play algorithm.

Source sentence
In this paper, we propose a numerical methodology for finding the closed-loop Nash equilibrium of stochastic delay differential games through deep learning. Our approach involves parameterizing the controls of each player using distinct recurrent neural networks.

Assumes 2

  • The initial path of the state is chosen so that the tax‑adjusted wealth Zi0 = Xi0 + aYi0 is positive. model
    Source sentence
    (43) We are assuming that the initial path Xi t = ζi(t) for t ≤0 is chosen so that Zi 0 = Xi 0 + aY i 0 > 0.
  • Each player’s control belongs to an admissible set Ai of closed‑loop, FX‑progressively measurable strategies. model
    Source sentence
    We know that in the continuous version, each player’s control lives in the admissible set Ai (2), which contains suitable FX t -progressively measurable strategies, i.e, closed-loop controls.

Data

  • synthetic benchmark problems with known closed‑form solutions synthetic unknown
    Source sentence
    This includes the new class of problems that we introduce and solve in this paper as well as the model introduced in [5] to study the systemic risk in bank lending. By considering both new and existing problems, we assess the accuracy of our proposed method using their closed-form solutions as benchmarks.

Costs

Not applicable The paper reports no trading or execution result.

Finds 2

  • The deep learning algorithm successfully approximates the true Nash equilibrium controls; the trajectories of the trained LSTM controls closely match those of the true controls after training. theoretical
    Source sentence
    The derivation, interpretation, and proofs for these problems are presented in Appendices A and B. From the numerical experiments we conducted, we have observed compelling evidence that our proposed algorithm successfully approximates the true Nash equilibrium controls. The trajectories of the trained controls closely match those of the true controls after training.
  • During training, the relative 2‑norm error between the predicted and true Nash equilibrium costs drops below 10⁻⁵ in some cases. performance
    Source sentence
    This is demonstrated by a decrease in the relative 2-norm error between these vectors throughout the training process, with relative errors reaching values lower than 10−5 in some cases.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 13 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.