Paper: arXiv 2507.01972

Authors: Hadi Keramati, Samaneh Jazayeri

Abstract

We present a reinforcement learning (RL)-driven framework for optimizing block-preconditioner sizes in iterative solvers used in portfolio optimization and option pricing. The covariance matrix in portfolio optimization or the discretization of differential operators in option pricing models lead to large linear systems of the form \(\mathbf{“A”}\textbf{“x”}=\textbf{“b”}\). Direct inversion of high-dimensional portfolio or fine-grid option pricing incurs a significant computational cost. Therefore, iterative methods are usually used for portfolios in real-world situations. Ill-conditioned systems, however, suffer from slow convergence. Traditional preconditioning techniques often require problem-specific parameter tuning. To overcome this limitation, we rely on RL to dynamically adjust the block-preconditioner sizes and accelerate iterative solver convergence. Evaluations on a suite of real-world portfolio optimization matrices demonstrate that our RL framework can be used to adjust preconditioning and significantly accelerate convergence and reduce computational cost. The proposed accelerated solver supports faster decision-making in dynamic portfolio allocation and real-time option pricing.

Complexity vs Empirical Score

  • Math Complexity: 7.5/10
  • Empirical Rigor: 6.0/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper employs advanced mathematical formulations including KKT conditions, finite difference discretizations of PDEs, and iterative solvers like FGMRES, indicating high math complexity. While it references real-world portfolio matrices and computational cost reductions, the lack of detailed backtest metrics, specific datasets, or implementation code places it in the moderate empirical rigor range.

Research Flowchart

  flowchart TD
  A["Research Goal: Accelerate iterative solvers<br>for portfolio & option pricing<br>using RL-driven preconditioning"] --> B["Methodology: RL Agent for<br>dynamic block-preconditioner sizing"]
  B --> C["Data Inputs: Real-world portfolio matrices &<br>option pricing discretization systems"]
  C --> D["Computational Process: RL-controlled<br>iterative solver with adaptive preconditioning"]
  D --> E{"Key Outcomes"}
  E --> F["Accelerated convergence of solvers"]
  E --> G["Reduced computational cost & time"]
  E --> H["Enhanced decision-making speed<br>for portfolio allocation & option pricing"]

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 reinforcement learning framework that dynamically adjusts block‑preconditioner sizes to accelerate iterative solvers for portfolio optimization and option pricing linear systems.

Source sentence
We present a reinforcement learning (RL)-driven framework for optimizing block-preconditioner sizes in iterative solvers used in portfolio optimization and option pricing. To overcome this limitation, we rely on RL to dynamically adjust the block-preconditioner sizes and accelerate iterative solver convergence.

Assumes 5

  • The problems can be expressed as large linear systems Ax = b derived from covariance matrices or discretized differential operators. model
    Source sentence
    The covariance matrix in portfolio optimization or the discretization of differential operators in option pricing models lead to large linear systems of the form $\mathbf{A}\textbf{x}=\textbf{b}$.
  • Iterative Krylov solvers (e.g., GMRES/FGMRES) are used to solve these linear systems. execution
    Source sentence
    Therefore, iterative methods are usually used for portfolios in real-world situations. We rely on the GMRES algorithm that begins by selecting an initial vector v1 and then uses the Arnoldi process to generate an orthonormal basis {v1, v2, . . . , vk} and an upper Hessenberg matrix Hk.
  • The block‑preconditioner size can be changed at each iteration and applied adaptively. other
    Source sentence
    The matrix A is split into smaller submatrices of size k × k, resulting in an approximate block diagonal matrix M: M = diag(A1, A2, . . . , Ak) (16) where Ai ∈Rs×s are the blocks of A. 4 Published as a Workshop paper at Advances in Financial AI Workshop (ICLR 2025) 3.2 BLOCK PRECONDITIONER Since we are changing the size of blocks of preconditioner in each iteration, we use Flexible GMRES (FGMRES). At each iteration,
  • A reinforcement learning agent can observe residuals and select an integer block size as its action. model
    Source sentence
    At each iteration, the agent observes the residuals of the linear system and, using its current policy, determines an optimal block size to apply the block-preconditioner.
Show 1 more
  • Proximal Policy Optimization (PPO) can learn an effective policy over multiple episodes. model
    Source sentence
    Over multiple episodes, the PPO agent learns a policy that effectively configures the preconditioning process to optimize the overall convergence rate.

Data

  • Real‑world portfolio optimization matrices benchmark unknown
    Source sentence
    Evaluations on a suite of real-world portfolio optimization matrices demonstrate that our RL framework can be used to adjust preconditioning and significantly accelerate convergence and reduce computational cost.
  • Option pricing coefficient matrices (e.g., size 2000) synthetic unknown
    Source sentence
    Figure 5 illustrates the convergence of the option pricing system for a coefficient matrix A of size 2000 at two different densities. In both cases, a clear speedup in obtaining the solution and achieving low residual values is observed across the different matrix sizes and densities.

Costs

Not applicable The paper reports no trading or execution result.

Finds 5

  • The RL‑based preconditioner significantly accelerates convergence and reduces computational cost for portfolio optimization problems. performance
    Source sentence
    Evaluations on a suite of real-world portfolio optimization matrices demonstrate that our RL framework can be used to adjust preconditioning and significantly accelerate convergence and reduce computational cost.
  • Across different matrix sizes and densities, a clear speedup and lower residual values are observed. performance
    Source sentence
    In both cases, a clear speedup in obtaining the solution and achieving low residual values is observed across the different matrix sizes and densities.
  • Using the RL agent can reduce the number of iterations required for option pricing to as few as two. performance
    Source sentence
    Although the computational cost of training the RL agent is nontrivial, the results demonstrate that using an RL agent can significantly reduce the number of iterations required to solve an option pricing problem , in some cases, to as few as two iterations.
  • Accelerated convergence is demonstrated for both portfolio optimization and option pricing across various matrix sizes and densities. performance
    Source sentence
    Ex-periments demonstrate accelerated convergence across various matrix sizes and densities for both portfolio optimization and option pricing.
  • The accelerated solver supports faster decision‑making in dynamic portfolio allocation and real‑time option pricing. other
    Source sentence
    The proposed accelerated solver supports faster decision-making in dynamic portfolio allocation and real-time option pricing.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 19 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 (32)

  1. Scaling Learning Algorithms Towards AI (2007)
  2. A Fast Learning Algorithm for Deep Belief Nets (2006)
  3. Deep learning (2016)
  4. Th 'e orie de la sp 'e culation (1900)
  5. The pricing of options and corporate liabilities (1973)
  6. Option pricing when underlying stock returns are discontinuous (1976)
  7. A Krylov subspace approach to large portfolio optimization (2012)
  8. Numerical methods and optimization in finance (2019)
and 24 more
  1. Iterative analysis (1962)
  2. Numerical stability of GMRES (1995)
  3. Any Nonincreasing Convergence Curve Is Possible for GMRES (1996)
  4. GMRES: A Generalized Minimal Residual Algorithm for Solving Nonsymmetric Linear Systems (1986)
  5. A Flexible Inner-Outer Preconditioned GMRES Algorithm (1993)
  6. Preconditioning Techniques for Large Linear Systems: A Survey (2002)
  7. Iterative Methods for Sparse Linear Systems (2003)
  8. Numerical Linear Algebra (2022)
  9. Graph Neural Preconditioners for Iterative Solutions of Sparse Linear Systems (2024) — link
  10. A Machine Learning Framework for Solving High-Dimensional Mean Field Game and Mean Field Control Problems (2018)
  11. Proximal policy optimization algorithms (2017) — link
  12. Proximal Policy Optimization Algorithms (2017) — link
  13. Numerical linear algebra (2022)
  14. An iterative approach to ill-conditioned optimal portfolio selection (2020)
  15. Numerical Methods for Pricing Multi-Asset Options (2017)
  16. Iterative methods for option pricing in Merton’s Jump diffusion model
  17. GPU based sparse grid technique for solving multidimensional options pricing PDEs (2009)
  18. Numerical Optimisation Problems in Finance (2017)
  19. Option pricing and portfolio optimization: modern methods of financial mathematics (2001)
  20. Efficient numerical methods for pricing American options under stochastic volatility (2008)
  21. Improved numerical solution of multi-asset option pricing problem: A localized RBF-FD approach (2019)
  22. Pricing multi-asset American options: A finite element method-of-lines with smooth penalty (2007)
  23. Compact finite difference method for American option pricing (2007)
  24. The University of Florida sparse matrix collection (2011)

0 of 32 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.