Neural Portfolio: Methods & Insights

• Neural Portfolio is a framework where neural networks parameterize key financial constructs—such as allocation weights and risk budgets—while embedding admissibility constraints.
• These methods leverage structured outputs, ranging from direct allocation to functionally generated portfolios, to address both construction and dynamic control challenges.
• Empirical findings indicate improved risk-adjusted returns compared to traditional methods, though challenges like transaction costs and scalability persist.

Searching arXiv for recent and foundational papers on neural portfolio methods. arXiv query: "neural portfolio optimization end-to-end portfolio functionally generated portfolio arXiv" Neural portfolio denotes a class of portfolio methods in which a neural network parameterizes a portfolio-relevant object—most commonly allocation weights, but also risk budgets, generating functions, execution controls, continuation values, hedge portfolios, or conditional valuation operators—and is trained against a financial objective such as relative wealth, Sharpe ratio, Omega ratio, Conditional Value-at-Risk, tracking difference, or conditional loss valuation (Ni et al., 2023, Uysal et al., 2021, Monoyios et al., 24 Jun 2025). In the literature considered here, the expression does not identify a single standardized model. This suggests a broader research program in which differentiable function approximation is coupled to portfolio constraints, stochastic control, optimization layers, and risk measurement.

1. Conceptual scope and taxonomy

The literature uses “Neural Portfolio” in several non-equivalent senses. In some papers, the network directly outputs portfolio weights and is trained end-to-end on portfolio performance. In others, the network outputs an intermediate object with financial structure: a risk-budget vector later converted into allocations by an implicit optimization layer, a concave generating function for a functionally generated portfolio, a leverage-feasible control in a continuous-time control problem, or a continuation-value surface for portfolio-level valuation (Fernandes et al., 16 May 2026, Uysal et al., 2021, Monoyios et al., 24 Jun 2025, Rossi et al., 2020).

Family	Learned object	Representative papers
Direct allocation	Portfolio weights $w_t$	(Fernandes et al., 16 May 2026, Nguyen et al., 2023, Lin et al., 29 Sep 2025)
Structured allocation	Risk budgets, generating functions, leverage-feasible controls	(Uysal et al., 2021, Monoyios et al., 24 Jun 2025, Ni et al., 2023)
Valuation and risk	Continuation values, conditional loss functions	(Rossi et al., 2020, Cheridito et al., 2021)
Execution and hedging	Selling schedules, compressed hedge portfolios	(Li et al., 2023, Dhandapani et al., 2024)

A useful distinction is between portfolio construction and portfolio valuation. The former includes end-to-end allocation, risk budgeting, functionally generated portfolios, and continuous-time dynamic portfolio choice. The latter includes future profit-and-loss distribution estimation, Value-at-Risk and Expected Shortfall estimation, portfolio compression, and exposure replication. A second distinction separates model-free architectures from model-based architectures. The model-free approach is described as a black-box in “End-to-End Risk Budgeting Portfolio Optimization with Neural Networks” (Uysal et al., 2021), whereas model-based designs embed financial structure directly into the network, for example through softmax budget constraints, convex optimization layers, Input Convex Neural Networks, or Pontryagin conditions (Uysal et al., 2021, Monoyios et al., 24 Jun 2025, Huh et al., 22 Jan 2025).

A common misconception is that “neural portfolio” necessarily means predicting returns and then applying a conventional optimizer. Several papers explicitly reject that interpretation. “Financially Guided Deep Portfolio Optimization” and “Decision-informed Neural Networks with LLM Integration for Portfolio Optimization” argue that predict-then-optimize suffers from a prediction–decision mismatch, while the parsimonious neural-network approach of (Staden et al., 2023) and the direct-allocation frameworks of (Nguyen et al., 2023) and (Lin et al., 29 Sep 2025) treat the decision rule itself as the learned object.

2. Neural parameterizations and admissibility constraints

The most direct parameterization maps historical features to allocation weights subject to long-only and full-investment constraints. In “Financially Guided Deep Portfolio Optimization,” the network outputs weights $w_t \in \mathbb{R}^N$ with $w_{t,i}\ge 0$ and $\sum_i w_{t,i}=1$, and all candidate architectures end with a softmax layer (Fernandes et al., 16 May 2026). “Cryptocurrency Portfolio Optimization by Neural Networks” uses a one-layer LSTM followed by a fully connected layer with Softmax output, again using the architecture itself to enforce nonnegativity and the budget constraint (Nguyen et al., 2023). “End-to-End Risk Budgeting Portfolio Optimization with Neural Networks” uses Softmax both in the model-free output layer and in the model-based risk budget layer (Uysal et al., 2021).

A second line of work does not learn weights directly. In the model-based risk budgeting framework, the network learns a target risk budget vector $b$, and a differentiable convex optimization layer implemented through CvxpyLayer solves the risk budgeting problem and returns the corresponding allocation (Uysal et al., 2021). In the Neural FGP framework, the network learns a generating function $G_\theta(\cdot)$, and the portfolio weights are then recovered from the classical functionally generated portfolio formula

$$\pi_{\theta,i}(x) = \left( \frac{\partial \log G_\theta(x)}{\partial x_i} + 1 - \sum_{j=1}^n x_j \frac{\partial \log G_\theta(x)}{\partial x_j} \right)x_i .$$

The architecture uses an Input Convex Neural Network with $G_\theta(x):=-f(x)$, where $f$ is convex, so $G_\theta$ is concave by construction (Monoyios et al., 24 Jun 2025).

Continuous-time portfolio choice introduces a different admissibility problem: the network must generate controls that remain feasible under leverage, borrowing, or consumption constraints. “Neural Network Approach to Portfolio Optimization with Leverage Constraints” proposes a leverage-feasible neural network (LFNN) that converts the original constrained optimization problem into an unconstrained optimization problem that is computationally feasible with standard optimization methods, and establishes mathematically that the LFNN approximation can yield a solution arbitrarily close to the original bounded-leverage control problem (Ni et al., 2023). In “Pontryagin-Guided Deep Learning for Large-Scale Constrained Dynamic Portfolio Choice,” the investment and consumption controls are neural networks, with a softplus output for the consumption network to enforce $w_t \in \mathbb{R}^N$0, while no-short-selling, borrowing limits, and consumption bounds are described as compatible with the framework by modifying the forward SDE, the Hamiltonian, or the network output constraints (Huh et al., 22 Jan 2025).

These designs show that neural portfolio methods are often less about unrestricted function approximation than about embedding admissibility into the parameterization. This suggests that financial structure is used not only for interpretability but also for numerical tractability and stability.

3. Objective functions and training criteria

The defining feature of many neural portfolio methods is that the training objective is a portfolio objective rather than a forecasting loss. “Financially Guided Deep Portfolio Optimization” turns Sharpe ratio, Omega ratio, Conditional Value-at-Risk, and Risk Parity into smooth differentiable surrogates and minimizes composite losses such as Sharpe-CVaR-RP and Omega-CVaR-RP (Fernandes et al., 16 May 2026). “Cryptocurrency Portfolio Optimization by Neural Networks” trains on the negative Sharpe ratio and augments it with a neutrality regularization term designed to reduce the network’s bias toward one asset and keep the portfolio close to a minimum-variance-like position (Nguyen et al., 2023). “From Headlines to Holdings” likewise uses a negative Sharpe-ratio-based loss for a daily allocation network combining LSTM, Graph Attention Networks, and news sentiment (Lin et al., 29 Sep 2025).

Other objectives are benchmark-relative or pathwise rather than absolute. Neural FGP maximizes terminal log relative wealth,

$w_t \in \mathbb{R}^N$1

and interprets learning as the data-driven choice of the generating function $w_t \in \mathbb{R}^N$2 of a functionally generated portfolio (Monoyios et al., 24 Jun 2025). The leverage-constrained high-inflation study formulates an optimal control problem to outperform a benchmark portfolio throughout the investment horizon under the cumulative quadratic tracking difference objective (Ni et al., 2023). The parsimonious dynamic portfolio framework handles mean-variance, mean-CVaR, quadratic target, one-sided quadratic loss, and mean semi-variance objectives in a single policy-learning setup (Staden et al., 2023).

A central controversy concerns whether minimizing prediction error suffices for good portfolio decisions. “Decision-informed Neural Networks with LLM Integration for Portfolio Optimization” argues both theoretically and empirically that minimizing the prediction error alone leads to suboptimal portfolio decisions, and trains the forecasting model through a differentiable mean–variance optimization layer with a hybrid loss

$w_t \in \mathbb{R}^N$3

Its decision loss is portfolio regret relative to an oracle portfolio, not a pure forecasting criterion (Hwang et al., 2 Feb 2025). The end-to-end risk budgeting paper makes the same criticism under the label “error maximization,” arguing that the usual two-step procedure can amplify intermediate estimation errors into poor allocations (Uysal et al., 2021).

Not all neural portfolio methods are end-to-end allocators. “ETF Portfolio Construction via Neural Network trained on Financial Statement Data” uses stock-level cross-entropy classification to predict whether a stock will go “up” or “down” over the next quarter, then aggregates stock scores through the ETF portfolio deposit file and forms an equal-weight portfolio from the top-ranked funds (Lee et al., 2022). In that setting, the neural network contributes ranking rather than direct weight optimization. This suggests that the term “neural portfolio” includes both end-to-end decision learning and neural scoring layers embedded in simpler portfolio rules.

4. Dynamic portfolio choice, stochastic control, and execution

A major branch of the literature treats neural portfolios as stochastic control problems. “A parsimonious neural network approach to solve portfolio optimization problems without using dynamic programming” parameterizes the policy as a single feedforward neural network $w_t \in \mathbb{R}^N$4 mapping time and state to portfolio weights, with the same network reused at every rebalance time. The paper emphasizes that the number of parameters remains independent of the number of rebalancing events, avoids the computation of high-dimensional conditional expectations, and proves convergence of the numerical solution to the theoretical optimal solution under general conditions (Staden et al., 2023). It also distinguishes time-consistent solutions for DP-separable objectives from pre-commitment solutions for non-separable objectives such as mean-variance and mean-CVaR.

Pontryagin-based methods pursue the same dynamic control agenda through first-order optimality conditions rather than backward dynamic programming. “Pontryagin-Guided Deep Learning for Large-Scale Constrained Dynamic Portfolio Choice” studies a continuous-time multi-asset Merton consumption–investment problem with neural-network controls $w_t \in \mathbb{R}^N$5 and $w_t \in \mathbb{R}^N$6, enforces Pontryagin’s Maximum Principle conditions during training, and reports scalability to $w_t \in \mathbb{R}^N$7 risky assets (Huh et al., 22 Jan 2025). The strongest empirical variant is PG-DPO-OneShot, which uses adjoint information to directly generate Pontryagin controls at test time.

The leverage-constrained high-inflation paper (Ni et al., 2023) also belongs to continuous-time stochastic control, but with a different emphasis: benchmark outperformance under a bounded leverage limit in a jump-diffusion high-inflation regime. Its contribution combines a closed-form unconstrained strategy under continuous trading and no bankruptcy with an LFNN approximation for realistic constrained trading.

Portfolio execution presents a related but distinct control problem. “Optimal Portfolio Execution in a Regime-switching Market with Non-linear Impact Costs” proposes a four-step numerical framework: approximate orthogonal portfolios from estimated impact matrices, solve each approximated orthogonal portfolio by dynamic program, pre-train a neural network on the resulting schedule, then fine-tune the network on the original coupled trading model (Li et al., 2023). In this setting, the neural network is not replacing dynamic programming outright; it is refining an approximate dynamic-programming solution and correcting the decomposition error caused by approximate orthogonalization.

These papers collectively show that neural portfolios need not be static allocators on a fixed feature vector. They can be feedback controls, consumption–investment policies, or liquidation schedules defined over continuous or discrete time.

5. Portfolio-level valuation, dependence, and risk management

Another important use of neural networks in portfolio problems is not allocation but portfolio-level valuation under nonlinear dependence. “Deep learning Profit & Loss” adapts Least Squares Monte Carlo by replacing polynomial regression of continuation values with a feed-forward neural network. The network is trained backward in time and has multiple outputs, one per portfolio component, so a single Monte Carlo simulation can interpolate every single asset in the portfolio while preserving strong dependence among contingent claims written on the same underlying (Rossi et al., 2020). The result is a neural-network version of LSMC for reconstructing the future P&L distribution of a multi-asset portfolio and deriving VaR, Expected Shortfall, and exposure profiles.

“Assessing asset-liability risk with neural networks” addresses a related conditional valuation problem. It approximates the conditional loss function

$w_t \in \mathbb{R}^N$8

with a feedforward neural network trained by mean-squared error, then computes Value-at-Risk and Expected Shortfall from Monte Carlo samples of the approximated loss $w_t \in \mathbb{R}^N$9. The paper supplements this with importance sampling and residual-based backtesting criteria for the defining conditional-expectation property (Cheridito et al., 2021). In this literature, the “portfolio” is a portfolio of assets and liabilities rather than an allocation vector over traded assets.

Sensitivity-based diversification offers a different portfolio-level representation. “Portfolio Optimization based on Neural Networks Sensitivities from Assets Dynamics respect Common Drivers” fits a neural network for each asset using common drivers, computes sensitivities by Automatic Adjoint Differentiation, embeds assets in a sensitivity space, and defines a Sensitivity matrix as the pairwise distance matrix between sensitivity vectors (Dominguez, 2022). Hierarchical clustering on this matrix yields Hierarchical Sensitivity Parity, an adaptation of Hierarchical Risk Parity in which diversification is driven by common-driver sensitivities rather than correlations alone.

Compression and static hedging constitute yet another portfolio-level application. “Neural Networks for Portfolio-Level Risk Management” designs an interpretable neural network in which each hidden unit corresponds to one option in a compressed portfolio, the hidden bias is the option strike, and the output-layer weight is the portfolio weight. The compressed portfolio is trained to approximate a larger European-option target portfolio across future risk horizons, and the paper reports close alignment of exposure distributions, Expected Exposure, Potential Future Exposure, and portfolio Greeks, together with lower standardized counterparty credit risk capital requirement for the compressed portfolio (Dhandapani et al., 2024).

These studies show that a neural portfolio can be a valuation operator, a dependence representation, or a static hedge. The common thread is portfolio-level function approximation under nonlinear payoffs and strong cross-asset dependence.

6. Empirical findings, limitations, and research directions

Empirical results vary widely because the underlying tasks differ. In benchmark-relative dynamic allocation under high inflation, the LFNN strategy is reported to consistently outperform the passive benchmark strategy by about 200 bps median annualized return, with a greater than 90% probability of outperforming the benchmark at the end of the investment horizon (Ni et al., 2023). In end-to-end risk budgeting on seven ETFs, the model-based Sharpe-trained portfolio achieves an out-of-sample Sharpe ratio of 1.16 for 2017–2021, and the gated end-to-end with filter boosts that figure to 1.24, compared with 0.79 for nominal risk parity and 0.83 for equal-weight fix-mix (Uysal et al., 2021). Neural FGP reports average terminal log relative return $w_{t,i}\ge 0$0 on real data, versus $w_{t,i}\ge 0$1 for equally weighted and smaller values for diversity-weighted benchmarks (Monoyios et al., 24 Jun 2025). A joint return-and-risk framework on ten large-cap U.S. equities reports annual return $w_{t,i}\ge 0$2 and Sharpe ratio $w_{t,i}\ge 0$3, outperforming equal weight and historical mean-variance benchmarks over 2020–2024 (Park, 9 Mar 2026).

End-to-end objective design also produces stronger results in difficult regimes. “Financially Guided Deep Portfolio Optimization” evaluates 50 S&P 500 stocks from 2007 to 2023 with quarterly rebalancing and bid-ask-spread transaction costs; on the 2022–2023 test period, the best model—AttentionLSTM with the Omega-CVaR-RiskParity loss—achieves annualized Sharpe 0.29 and total compounded return $w_{t,i}\ge 0$4, while the S&P 500 delivers $w_{t,i}\ge 0$5 total return and annualized Sharpe $w_{t,i}\ge 0$6, with almost no increase in CVaR (Fernandes et al., 16 May 2026). “Decision-informed Neural Networks with LLM Integration for Portfolio Optimization” reports on S&P100 and DOW30 that the decision-informed architecture exceeds a range of Transformer-style and LLM-based baselines on return, Sharpe, Sortino, terminal wealth, and return-over-VaR, while gradient analyses indicate that the model prioritizes the assets most crucial to decision making (Hwang et al., 2 Feb 2025).

At the same time, the literature is explicit about limitations. Several studies work on small universes: nine U.S. stocks in the price–graph–sentiment architecture (Lin et al., 29 Sep 2025), a pair of cryptocurrency derivative assets in the BLVT hedging study (Nguyen et al., 2023), or ten large-cap U.S. equities in the joint return-and-risk model (Park, 9 Mar 2026). Some omit transaction costs (Park, 9 Mar 2026), assume zero fees (Gao et al., 2020, Lin et al., 29 Sep 2025), or rely on penalty methods that can compromise feasibility. “Large-scale portfolio optimization with variational neural annealing” reports near-optimal solutions comparable to Mosek on portfolios with 478, 857, and 2008 assets, but also notes that validity can drop to $w_{t,i}\ge 0$7 under the penalty-only approach on Russell 3000 before improving with a Lagrange-multiplier modification (Ranabhat et al., 9 Jul 2025). The spiking neural network cross-market study claims higher cumulative return, better rolling Sharpe, lower concentration, and improved computational efficiency relative to an ANN benchmark, but it also acknowledges more training epochs, dependence on hierarchical clustering, and the absence of real neuromorphic deployment (Mohan et al., 1 Oct 2025).

A further limitation is conceptual rather than computational. High predictive accuracy is not the same as high decision quality, and the literature increasingly treats this as a core design problem rather than a minor implementation detail. Another is that time inconsistency remains inherent for objectives such as mean-variance, mean-CVaR, and mean semi-variance; the parsimonious dynamic framework explicitly treats the resulting controls as pre-commitment solutions when dynamic programming separability is absent (Staden et al., 2023).

The present state of the field therefore supports two broad conclusions. First, neural portfolio research has moved beyond “predict returns with a neural network.” It now includes differentiable allocators, optimization layers, stochastic-control policies, functionally generated portfolios, valuation surrogates, sensitivity-space diversification, and interpretable static hedge compressors. Second, the strongest methods usually embed substantial financial structure—budget constraints, leverage feasibility, concavity, risk budgeting, pathwise decomposition, or optimal-control conditions—rather than relying on unconstrained black-box mapping alone. This suggests that the most durable notion of a neural portfolio is not a particular architecture, but a design principle: learn a portfolio object with neural networks while preserving enough financial structure for admissibility, interpretability, and out-of-sample robustness.