Paper: arXiv 2411.09899
Authors: Yaacov Kopeliovich, Michael Pokojovy
Abstract
With the recent advancements in machine learning (ML), artificial neural networks (ANN) are starting to play an increasingly important role in quantitative finance. Dynamic portfolio optimization is among many problems that have significantly benefited from a wider adoption of deep learning (DL). While most existing research has primarily focused on how DL can alleviate the curse of dimensionality when solving the Hamilton-Jacobi-Bellman (HJB) equation, some very recent developments propose to forego derivation and solution of HJB in favor of empirical utility maximization over dynamic allocation strategies expressed through ANN. In addition to being simple and transparent, this approach is universally applicable, as it is essentially agnostic about market dynamics. To showcase the method, we apply it to optimal portfolio allocation between a cash account and the S&P 500 index modeled using geometric Brownian motion or the Heston model. In both cases, the results are demonstrated to be on par with those under the theoretical optimal weights assuming isoelastic utility and real-time rebalancing. A set of R codes for a broad class of stochastic volatility models are provided as a supplement.
Complexity vs Empirical Score
- Math Complexity: 7.5/10
- Empirical Rigor: 7.0/10
- Quadrant: Holy Grail — high math complexity, high empirical rigor
Why this score: The paper employs advanced mathematics including stochastic differential equations, Hamilton-Jacobi-Bellman equations, and neural network theory, yet provides a backtest-ready framework with R code and validation on S&P 500 and VIX data.
Research Flowchart
flowchart TD
A["Research Goal: <br>Dynamic Portfolio Optimization with ANN"] --> B["Key Methodology: <br>Empirical Utility Maximization via ANN"]
B --> C["Data & Inputs: <br>S&P 500 & Cash via GBM/Heston Models"]
C --> D["Computational Process: <br>Train ANN to Maximize Cumulative Utility"]
D --> E{"Key Findings & Outcomes"}
E --> F["ANN strategies match <br>theoretical optimal weights"]
E --> G["Method is model-agnostic <br>and avoids solving HJB"]
E --> H["Code provided for <br>stochastic volatility models"]
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 deep‑learning framework that uses empirical utility maximization with artificial neural networks to obtain optimal dynamic portfolio allocation without deriving or solving the Hamilton‑Jacobi‑Bellman equation.
Source sentence
While most existing research has primarily focused on how DL can alleviate the curse of dimensionality when solving the Hamilton-Jacobi-Bellman (HJB) equation, some very recent developments propose to forego derivation and solution of HJB in favor of empirical utility maximization over dynamic allocation strategies expressed through ANN. In other words, we consider the Merton problem and attempt to apply AI tools
Assumes 6
- The risky asset follows either a geometric Brownian motion or a Heston stochastic volatility model. model
Source sentence
To showcase the method, we apply it to optimal portfolio allocation between a cash account and the S&P 500 index modeled using geometric Brownian motion or the Heston model. Examples We present two application examples to showcase how the DL approach from Section 3 can be applied to optimal portfolio problem (1)–(7) when the risky asset is modeled using geometric Brownian motion (Section 4.1) or Heston model
- The risk‑free rate is constant at 5% per annum. market
Source sentence
The risk-free rate was set at 5% p.a.
- Initial wealth is set to $1.0. other
Source sentence
The initial wealth was set at $1.0.
- Three years of adjusted daily closing prices (1/1/2021–12/31/2023) are used for calibration of the market models. data
Source sentence
Three years’ worth of adjusted daily closing prices (1/1/2021–12/31/2023) were used for calibration purposes3 as detailed below.
Show 2 more
- Synthetic market paths can be generated arbitrarily many times for training the ANN. statistical
Source sentence
(21) 3Note that the number of data points (753 here) used for model calibration does not limit our ability to simulate arbitrarily many synthetic paths to train the ANN model.
- Training, test and validation splits are used to avoid over‑training. execution
Source sentence
However, utilization of data splitting into training, test and validation segments is crucial to protect against overtraining.
Data
- S&P 500 adjusted daily closing prices market-prices public
Source sentence
In both situations, the S&P 500 index (GSPC ticker) was selected as risky asset, while VIX index was used to calibrate Heston volatility process. Three years’ worth of adjusted daily closing prices (1/1/2021–12/31/2023) were used for calibration purposes3 as detailed below.
- VIX index data market-prices public
Source sentence
In both situations, the S&P 500 index (GSPC ticker) was selected as risky asset, while VIX index was used to calibrate Heston volatility process.
- Synthetic market paths (GBM and Heston) synthetic unknown
Source sentence
To avoid overtraining, minibatches were re-simulated at each step. (21) 3Note that the number of data points (753 here) used for model calibration does not limit our ability to simulate arbitrarily many synthetic paths to train the ANN model.
Costs
Not modeled
Finds 5
- For both GBM and Heston models, the ANN‑based allocation yields expected utility on par with the analytic optimal weights. performance
Source sentence
Good agreement with the analytic optimum can 8 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 1/ η Stock Weight Theoretical ANN Figure 2: Merton ratio vs 1/η. be observed. Again, the expected utility under ANN was on par with that under the analytic weight from Equation (22) over the range of η’s considered.
- Training the ANN required 27,000 (GBM) to 45,000 (Heston) gradient evaluations, far fewer than the 2.56 million simulations used in prior work. performance
Source sentence
While van Staden et al. (2023) required 2.56 million replications, 27,000 (for GBM) to 45,000 (for Heston) gradient evaluations were sufficient in our simulations.
- The total training time was less than 6 hours. performance
Source sentence
The training time was less than 6 hrs.
- With only 3–5 hidden neurons in a single hidden layer, the ANN reproduces analytic portfolio weights and utility values consistently. performance
Source sentence
With only 3-5 hidden neurons and a single hidden layer, the ANN results appeared to be consistent with the analytic ones both in terms of portfolio weights and expected utility values.
- All algorithms were implemented in plain R code without third‑party deep‑learning libraries or GPU acceleration. other
Source sentence
All algorithm were implemented in plain R code and run in RStudio under 64-bit Windows 11 Pro OS on a Dell Precision 3581 laptop with 13th Gen Intel(R) Core(TM) i7-13800H 2.50 GHz CPU and 32 GB RAM. No third-party DL libraries or GPU acceleration were employed. 4.1.
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 28 related papers). Coverage is partial, so this is neither replication nor refutation.
2 papers in the archive
Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.