Paper: arXiv 2506.19715

Authors: Michael Monoyios, Olivia Pricilia

Abstract

We introduce a novel neural-network-based approach to learning the generating function \(G(\cdot)\) of a functionally generated portfolio (FGP) from synthetic or real market data. In the neural network setting, the generating function is represented as \(G_θ(\cdot)\), where \(θ\) is an iterable neural network parameter vector, and \(G_θ(\cdot)\) is trained to maximise investment return relative to the market portfolio. We compare the performance of the Neural FGP approach against classical FGP benchmarks. FGPs provide a robust alternative to classical portfolio optimisation by bypassing the need to estimate drifts or covariances. The neural FGP framework extends this by introducing flexibility in the design of the generating function, enabling it to learn from market dynamics while preserving self-financing and pathwise decomposition properties.

Complexity vs Empirical Score

  • Math Complexity: 8.5/10
  • Empirical Rigor: 4.0/10
  • Quadrant: Lab Rats — high math complexity, limited empirical testing

Why this score: The paper is heavily mathematical, featuring advanced stochastic calculus, partial differential equations, and pathwise decomposition theorems typical of Stochastic Portfolio Theory. However, it lacks specific backtesting results, performance metrics, or implementation details, relying instead on synthetic data and theoretical propositions rather than real-world empirical validation.

Research Flowchart

  flowchart TD
  A["Research Goal: Learn generating function G_θ(·)<br>to optimize portfolio return"] --> B["Neural FGP Methodology"]
  
  B --> C["Data Input: Synthetic/Real Market Data"]
  
  C --> D["Computational Process: Train G_θ(·)<br>via Maximizing Investment Return"]
  
  D --> E["Preserve Self-Financing &<br>Pathwise Decomposition Properties"]
  
  E --> F["Compare against Classical FGP Benchmarks"]
  
  F --> G["Key Findings: Neural FGP outperforms benchmarks<br>with flexible learning from market dynamics<br>while maintaining robustness without<br>drift/covariance estimation"]

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 neural-network-based approach to learn the generating function of a functionally generated portfolio directly from data.

Source sentence
We introduce a novel neural-network-based approach to learning the generating function G(·) of a functionally generated portfolio (FGP) from synthetic or real market data.

Assumes 6

  • The market satisfies the no unbounded profit with bounded risk (NUPBR) condition. market
    Source sentence
    SPT operates under a weaker assumption: no unbounded profit with bounded risk (NUPBR), which is equivalent to the existence of a non-empty set of local martingale deflators (LMDs).
  • Portfolio wealth processes are non‑negative for all times. model
    Source sentence
    We assume the admissibility condition P (V w,π(t) ≥0, ∀t ∈[0, T]) = 1.
  • The generating function G(·) is C² and strictly positive on an open neighbourhood of the simplex. model
    Source sentence
    A portfolio π is said to be functionally generated if there exists a C2, strictly positive function G : U
  • Synthetic data are generated as geometric Brownian motion (GBM) price paths with randomly sampled drifts and volatilities over 1000 days. data
    Source sentence
    For synthetic data, we generate geometric Brownian motion (GBM) paths for asset prices and normalise them to compute market weights.
Show 2 more
  • Real‑world data are daily adjusted closing prices fetched from Yahoo Finance for selected tickers over the past five years. data
    Source sentence
    For real data, we fetch daily adjusted closing prices for the selected tickers (e.g., AAPL, MSFT, GOOG, AMZN, META) over the past y = 5 years via Yahoo Finance, ending on Friday, 11 April 2025.
  • Training maximises logarithmic portfolio returns relative to the market over a 200‑day window. execution
    Source sentence
    The ICNN is trained via back-propagation to maximise logarithmic portfolio returns relative to the market over a window of 200 trading days.

Data

  • Synthetic GBM price paths synthetic unknown
    1000 days
    Source sentence
    For synthetic data, we generate geometric Brownian motion (GBM) paths for asset prices and normalise them to compute market weights.
  • Yahoo Finance daily adjusted closing prices · Yahoo Finance market-prices public
    past 5 years ending 11 April 2025 · AAPL, MSFT, GOOG, AMZN, META
    Source sentence
    For real data, we fetch daily adjusted closing prices for the selected tickers (e.g., AAPL, MSFT, GOOG, AMZN, META) over the past y = 5 years via Yahoo Finance, ending on Friday, 11 April 2025.

Recognised sources: Yahoo Finance

Costs

Not modeled A trading or execution result is reported without costs.

Finds 3

  • Neural FGP achieves an average logarithmic relative return of 0.078694 on synthetic data, outperforming equally‑weighted, market, and diversity‑weighted benchmarks. performance
    Source sentence
    FGP 0.078694 EWP 0.0092019 Market -9.79007e-07 DWP p = 0.3 0.00660325 DWP p = 0.5 0.00478323 DWP p = 0.8 0.0019446
  • Across all walk‑forward windows, the Neural FGP consistently outperforms the classical benchmark strategies in both synthetic and real‑world evaluations. performance
    Source sentence
    In both settings, the Neural FGP consistently outperforms all benchmark strategies across nearly all evaluation periods.
  • The Neural FGP framework preserves self‑financing, pathwise decomposition, and dependence on observable quantities while allowing flexible generating‑function design. theoretical
    Source sentence
    In our framework, the generating function G(·) is parameterised by a neural network trained on historical market data. This allows the portfolio to adapt to complex, possibly nonlinear features in the market weight dynamics, without sacrificing the interpretability and robustness of the original FGP formulation.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 17 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 (15)

  1. Ergodic robust maximization of asymptotic growth (2018) — link
  2. Input Convex Neural Networks (2016) — link
  3. Stochastic Portfolio Theory: A Machine Learning Approach (2016) — link
  4. On optimal arbitrage (2010) — link
  5. Robust maximization of asymptotic growth (2010) — link
  6. Diversity and relative arbitrage in equity markets (2008) — link
  7. The numéraire portfolio in semimartingale financial models (2007) — link
  8. Stochastic Portfolio Theory (2002) — link
and 7 more
  1. On the diversity of equity markets (1999) — link
  2. Portfolio Generating Functions (1998) — link
  3. A general version of the fundamental theorem of asset pricing (1994) — link
  4. ON THE REPRESENTATION OF SEMIMARTINGALES (1973) — link
  5. The exit measure of a supermartingale (1972) — link
  6. Probabilistic Aspects of Arbitrage (2010) — link
  7. Stochastic Portfolio Theory: an Overview (2009) — link

0 of 15 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.