Paper: arXiv 2511.07045

Authors: John Armstrong, Cristin Buescu, James Dalby, Rohan Hobbs

Abstract

We use a neural network to identify the optimal solution to a family of optimal investment problems, where the parameters determining an investor’s risk and consumption preferences are given as inputs to the neural network in addition to economic variables. This is used to develop a practical tool that can be used to explore how pension outcomes vary with preference parameters. We use a Black-Scholes economic model so that we may validate the accuracy of network using a classical and provably convergent numerical method developed using the duality approach.

Complexity vs Empirical Score

  • Math Complexity: 8.0/10
  • Empirical Rigor: 3.0/10
  • Quadrant: Lab Rats — high math complexity, limited empirical testing

Why this score: The paper employs advanced mathematics including stochastic differential equations, Itô’s lemma, exponential utility functions, and duality approaches for solving a high-dimensional optimization problem. However, it relies on a simplified Black-Scholes model for validation and focuses on theoretical numerical methods without presenting backtests, real-world data, or implementation-heavy empirical results.

Research Flowchart

  flowchart TD
  A["Research Goal:<br>Identify optimal pension investment solutions<br>for a family of preference parameters"] --> B["Data & Inputs:<br>Economic Model: Black-Scholes<br>Preferences: Risk & Consumption parameters"]
  
  B --> C1["ML Methodology:<br>Neural Network Training<br>using duality-based optimization"]
  B --> C2["Validation Methodology:<br>Classical Numerical Solver<br>for ground truth comparison"]
  
  C1 --> D["Computational Process:<br>Neural Network approximates<br>optimal investment strategy function"]
  C2 --> D
  
  D --> E["Key Findings:<br>1. Validated NN accuracy vs. classical method<br>2. Developed practical tool for exploring<br>pension outcomes vs. preferences<br>3. Generalizable family of 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 using a neural network to identify the optimal solution to a family of optimal pension investment problems, taking investor risk and consumption preference parameters as inputs alongside economic variables.

Source sentence
We use a neural network to identify the optimal solution to a family of optimal investment problems, where the parameters determining an investor's risk and consumption preferences are given as inputs to the neural network in addition to economic variables.

Assumes 6

  • The investment problem is set within a Black‑Scholes framework with a risky asset following geometric Brownian motion. market
    Source sentence
    1 Discrete-time investment problem We consider an optimal investment and consumption problem set within a Black-Scholes framework. The dynamics of the risky asset price, denoted S, follow a geometric Brownian motion described by dSt = µStdt + σSt dWt, where µ ∈R represents the drift, σ ∈R+ the volatility and W a standard Brownian motion.
  • Investment and consumption decisions are made at discrete time intervals. execution
    Source sentence
    Investment and consumption decisions are assumed to be made at discrete intervals, defined by the set T := {0, δt, 2δt, . . . , T} for some time-step δt and final time T. Between these times, investments are made following a fixed-weight strategy.
  • The risk‑free asset grows at a constant rate r. model
    Source sentence
    This proprtion is maintained throughout the interval [t, t+δt) with the remainder of the portfolio allocated to a risk-free asset, growing at a constant rate r.
  • Consumption and individual longevity risk are independent. statistical
    Source sentence
    In order to compute the gain function, we assume that consumption and individual longevity risk are independent.
Show 2 more
  • There is no systematic longevity risk; mortality follows given yearly probabilities ps. statistical
    Source sentence
    We also assume there is no systematic longevity risk and that the probability an individual dies in a given year s is given by ps.
  • The gain function is exponential, leading to high variance across parameter values. other
    Source sentence
    Since we employ an exponential gain func-tion, the loss values can vary significantly across differing parameter values. As a result, when averaging the loss function across scenarios, certain parameter combinations can dominate as they provide more extreme utility values with more extreme variance.

Data

  • Synthetic Gaussian increments for log‑wealth simulation synthetic unknown
    Source sentence
    For simplicity and simulation, we define ϵt := Wt+δt −Wt √ δt , (2) so ϵt is distributed according to standard normal. For this Black–Scholes model, we take the standard Gaussians ϵt in equation (2) as the input to the recurrent neural network that we train.

Costs

Not applicable The paper reports no trading or execution result.

Finds 5

  • The neural network’s optimal strategies match those obtained by a provably convergent classical algorithm, with expected utility within 1% of the benchmark for the default parameter set. theoretical
    Source sentence
    Furthermore, the expected utility obtained by the varying-parameter network is within 1% of a standard error of the value obtained by the fixed-parameter network for this default set of parameters.
  • Only three full iterations of the scaling‑network procedure were needed to achieve sufficient accuracy. theoretical
    Source sentence
    In practice, we found that only three full iterations of this process were necessary to solve our problem with sufficient accuracy.
  • The varying‑parameter network’s performance remains within 1% of the fixed‑parameter network for edge‑case parameter combinations. performance
    Source sentence
    The varying-parameter network was also within 1% under these parameters.
  • The two‑step iterative architecture improves the network’s ability to generalize across a wide range of preference parameters. performance
    Source sentence
    Figure 4a demonstrates the improved ability of the network to generalize across a wide range of parameters.
  • Higher risk aversion leads to lower risky‑asset allocation and narrower replacement‑ratio percentile spreads, while lower risk aversion results in higher consumption in retirement. other
    Source sentence
    Specifically, indi-viduals with higher risk aversion allocate a lower proportion to the risky asset, resulting in narrower spreads between replacement-ratio percentiles. In terms of consump-tion strategies, less risk-averse individuals tend to consume more, reflecting a reduced concern for the depletion of their funds.

Checked by others

Supported by

  • Equilibrium Investment with Random Risk Aversion: (Non-)uniqueness, Optimality, and Comparative Statics (2025)
    On the question: How does risk aversion influence the amount of risky‑asset investment?
    Compare the findings

    This paper: Higher risk aversion leads to lower risky‑asset allocation and narrower replacement‑ratio percentile spreads, while lower risk aversion results in higher consumption in retirement.

    That paper: A larger risk aversion under the reverse hazard rate order always leads to less risky investment.

    Both findings agree that higher risk aversion reduces risky‑asset allocation (A: lower risky‑asset share and narrower replacement‑ratio spreads; B: less risky investment).

  • Robust optimal consumption, investment and reinsurance for recursive preferences (2025)
    On the question: How does risk aversion influence optimal consumption and risky‑asset allocation decisions?
    Compare the findings

    This paper: Higher risk aversion leads to lower risky‑asset allocation and narrower replacement‑ratio percentile spreads, while lower risk aversion results in higher consumption in retirement.

    That paper: Optimal consumption decreases with higher risk aversion and higher elasticity of intertemporal substitution (EIS).

    Both findings agree that higher risk aversion reduces consumption and lowers risky‑asset allocation; B also notes narrower replacement‑ratio percentile spreads under higher risk aversion.

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 (81)

  1. X marks the bank: digitalisation and depositor runs (2025)
  2. Directed Scale-Free Graphs (2003) — link
  3. Pareto’s Law of Income Distribution: Evidence for Germany, the United Kingdom, and the United States (2005) — link
  4. Do wealth distributions follow power laws? Evidence from ‘rich lists’ (2014) — link
  5. The Double Pareto-Lognormal Distribution—A New Parametric Model for Size Distributions (2004) — link
  6. A comparison of the effectiveness of alternative DC and CDC designs in a UK market — → 2504.16892
  7. Optimal post-retirement investment and consumption under longevity risk in collective funds (2025) — link
  8. Demand–Deposit Contracts and the Probability of Bank Runs (2005) — link
and 73 more
  1. Bank Runs, Deposit Insurance, and Liquidity (1983)
  2. Unique Equilibrium in a Model of Self-Fulfilling Currency Attacks (1998) — link
  3. Global Games and Equilibrium Selection (1993) — link
  4. Utility Theory Front To Back — Inferring Utility From Agents’ Choices (2014) — link
  5. Importance sampling for option pricing with feedforward neural networks (2025) — link
  6. Social Media as a Bank Run Catalyst (2024) — link
  7. Statistical mechanics of complex networks (2002) — link
  8. The Structure and Function of Complex Networks (2003) — link
  9. On Random Graphs I (1959)
  10. Intergenerational cross-subsidies in UK collective defined contribution (CDC) funds — → 2411.13565
  11. Collectivised Post-Retirement Investment (2020) — link
  12. Collectivised Pension Investment with Exponential Kihlstrom–Mirman Preferences (2019) — link
  13. Optimal consumption and portfolio choice for pooled annuity funds (2008)
  14. An elementary approach to the Merton problem (2021)
  15. Theory of Games and Economic Behavior (1947)
  16. A Note on the Pure Theory of Consumer’s Behaviour (1938)
  17. The Construction of Utility Functions from Expenditure Data (1967)
  18. The Nonparametric Approach to Demand Analysis (1982)
  19. Risks for the Long Run: A Potential Resolution of Asset Pricing Puzzles (2004) — link
  20. Long-run risks and financial markets (2007) — link
  21. Explaining asset pricing puzzles associated with the 1987 market crash (2011) — link
  22. Optimal mutual insurance against systematic longevity risk — → 2410.07749
  23. The infinite-horizon investment-consumption problem for Epstein-Zin stochastic differential utility. II (2023) — link
  24. The infinite horizon investment-consumption problem for Epstein-Zin stochastic differential utility. I : Foundations (2023) — link
  25. Optimal consumption and investment with Epstein–Zin recursive utility (2017)
  26. King William’s Tontine: Why the Retirement Annuity of the Future Should Resemble its Past (2015)
  27. Optimal retirement income tontines (2015) — link
  28. Modern tontine with bequest: Innovation in pooled annuity products (2019) — link
  29. Bringing cost transparency to the life annuity market (2014) — link
  30. Lifetime Portfolio Selection under Uncertainty: The Continuous-Time Case (1969)
  31. Dynamic Programming (1957)
  32. Continuous-time Stochastic Control and Optimization with Financial Applications (2009)
  33. Learning a functional control for high-frequency finance (2022) — link
  34. Uncovering Utility Functions from Observed Outcomes (2025) — link
  35. Uncertain lifetime and intertemporal choice: risk aversion as a rationale for time discounting (2006)
  36. Temporal resolution of uncertainty and dynamic choice theory (1978)
  37. Classifying Markets up to Isomorphism (2018) — link
  38. Convex Analysis (2015)
  39. Optimization by vector space methods (1997)
  40. Grouping individual investment preferences in retirement savings: A cluster analysis of a USS members risk attitude survey (2020) — link
  41. Destabilisation of bank deposits across destinations: assessment and policy implications (2024) — link
  42. Understanding the Speed and Size of Bank Runs in Historical Comparison (2023) — link
  43. A Monetary History of the United States, 1867-1960 (1963)
  44. Securitized banking and the run on repo (2012) — link
  45. Deep neural networks algorithms for stochastic control problems on finite horizon: convergence analysis (2021)
  46. Deep learning approximation for stochastic control problems (2016) — link
  47. Recent developments in machine learning methods for stochastic control and games (2024) — link
  48. ILLIQUIDITY COMPONENT OF CREDIT RISK – THE 2015 LAWRENCE R. KLEIN LECTURE (2016) — link
  49. Strategic Complementarity, Fragility, and Regulation (2014) — link
  50. Coordination Failures and the Lender of Last Resort: Was Bagehot Right After All? (2004) — link
  51. Two centuries of systemic bank runs (2024) — link
  52. Deal for failed Silicon Valley carries record estimated cost to FDIC at $20B (2023) — link
  53. Does One Soros Make a Difference? A Theory of Currency Crises with Large and Small Traders (2004) — link
  54. Public Information, Private Information, and the Multiplicity of Equilibria in Coordination Games (2002) — link
  55. A General Theory of Equilibrium Selection in Games (1988) — link
  56. Machine-learning a family of solutions to an optimal pension investment problem (2025) — link
  57. Application of data clustering and machine learning in variable annuity valuation (2013) — link
  58. A data-driven neural network approach to optimal asset allocation for target based defined contribution pension plans (2019) — link
  59. A hybrid deep learning method for optimal insurance strategies: Algorithms and convergence analysis (2021) — link
  60. Bayesian CART models for insurance claims frequency — → 2303.01923
  61. Machine learning for retirement planning (2020)
  62. Heterogeneous Retirement Savings Strategy Selection with Reinforcement Learning (2023)
  63. A neural network approach to efficient valuation of large portfolios of variable annuities (2016) — link
  64. Handbook of Convergence Theorems for (Stochastic) Gradient Methods (2024) — link
  65. Risk aversion with many commodities (1974) — link
  66. Substitution, Risk Aversion, and the Temporal Behavior of Consumption and Asset Returns: A Theoretical Framework (1989) — link
  67. Efficient Per-Example Gradient Computations (2015) — link
  68. Tutorial on amortized optimization (2023)
  69. Deep hedging: continuous reinforcement learning for hedging of general portfolios across multiple risk aversions (2022) — link
  70. GradNorm : gradient normalization for adaptive loss balancing in deep multitask networks (2018)
  71. Learning values across many orders of magnitude (2016)
  72. A Hybrid Deep Reinforcement Learning Method for Insurance Portfolio Management (2026) — link
  73. Monte Carlo Methods in Financial Engineering (2004)

4 of 81 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.