Quantum-Enhanced Portfolio Optimization

• Quantum-enhanced portfolio optimization is a research area that applies quantum algorithms and hybrid quantum-classical methods to address complex asset allocation challenges.
• It involves diverse formulations including online, mean-variance, and discrete selection techniques, utilizing QUBO/Ising mappings, variational methods, and amplitude-based encodings.
• Practical implementations face challenges such as data-access bottlenecks, noise tolerance, and hardware limitations, prompting the integration of hybrid and post-processing strategies.

Quantum-enhanced portfolio optimization denotes a heterogeneous research area in which quantum algorithms, quantum-inspired heuristics, and hybrid quantum-classical workflows are used to solve portfolio construction problems that range from online adversarial allocation to discrete mean-variance selection, higher-moment optimization, Black-Litterman inference, ESG-constrained allocation, and graph-structured diversification. The literature does not describe a single canonical model. Instead, it spans oracle-based algorithms with regret guarantees, QUBO/Ising encodings for annealers and gate-model variational methods, sampling-based variational schemes that avoid diagonal-Hamiltonian overhead, and quantum-inspired metaheuristics for high-dimensional Sharpe-ratio optimization (Lim et al., 2022, Rebentrost et al., 2018, Uotila et al., 1 Sep 2025, Chen et al., 2023, Agliardi et al., 19 Aug 2025, Yu et al., 16 Jan 2026).

1. Financial problem classes and objective functions

A first major branch studies online portfolio optimization. In that setting, there are $T$ discrete trading rounds and $n$ assets; at each round $t$, the learner chooses a long-only portfolio vector $w^{(t)} \in \Delta^n$, the market reveals a nonnegative return vector, and wealth evolves multiplicatively:

$$S_{t+1} = S_t (w^{(t)} \cdot r^{(t)}), \qquad S_T = S_0 \prod_{t=1}^T (w^{(t)} \cdot r^{(t)}).$$

The objective is the normalized average log-return

$$L_T := \frac{1}{T} \sum_{t=1}^T \log(w^{(t)} \cdot r^{(t)}),$$

with regret measured against the best fixed portfolio in hindsight. The model in this line is explicitly long-only, normalized to the simplex, and can be online and potentially adversarial (Lim et al., 2022).

A second branch uses continuous mean-variance optimization in the Markowitz sense. One canonical form minimizes $w^T \Sigma w$ subject to $\mu^T w = \mu_p$ and $\mathbf{1}^T w = 1$, with efficient-frontier quantities

$$A = \mathbf{1}^T \Sigma^{-1} \mathbf{1},\quad B = \mathbf{1}^T \Sigma^{-1} \mu,\quad C = \mu^T \Sigma^{-1} \mu,\quad D = AC - B^2,$$

and optimal weights

$n$0

This formulation is used in quantum linear-system approaches, where short selling is allowed in the baseline formulation and the output of primary interest is often a quantum state proportional to the optimal portfolio (Rebentrost et al., 2018).

A third branch focuses on discrete portfolio selection, typically with binary decisions, integer holdings, or fixed lots. Representative formulations include

$n$1

for binary cardinality-constrained mean-variance selection, and

$n$2

for higher-order moment optimization with integer holdings and a capital budget. The latter explicitly incorporates skewness and kurtosis through

$n$3

and leads naturally to a higher-order unconstrained binary optimization (HUBO) rather than a QUBO (Innan et al., 28 Jul 2025, Uotila et al., 1 Sep 2025).

Other financially distinct formulations also appear. In Black-Litterman settings, the posterior mean

$n$4

is used inside a discrete mean-variance selection problem, with investor views predicted by quantum machine learning (Chen et al., 2023). In ESG-aware discrete Markowitz models, the optimization adds a Wasserstein-style ESG constraint

$n$5

rather than a simple linear ESG reward term (Catalano et al., 2024). In realistic ETF bond construction, the task becomes the minimization of class-level metric deviations under budget and guardrail constraints, then reduced to

$n$6

or to an unconstrained hinge-penalized objective evaluated directly from samples (Agliardi et al., 19 Aug 2025).

2. Quantum formulations, encodings, and state representations

The dominant representation for discrete finance problems is the QUBO/Ising mapping. With binary variables encoded as

$n$7

a quadratic objective becomes a diagonal cost Hamiltonian

$n$8

This mapping is used across penalty-based QAOA, SamplingVQE, annealing, and Black-Litterman-based discrete selection. In higher-moment models, the same substitution produces a diagonal Hamiltonian with 1-, 2-, 3-, and 4-body Pauli-$n$9 strings because kurtosis induces quartic terms (Innan et al., 28 Jul 2025, Uotila et al., 1 Sep 2025, Chen et al., 2023).

A separate formulation avoids explicit diagonal-Hamiltonian compilation. In sampling-based variational portfolio construction for ETF bond design, the objective is evaluated classically on measured bitstrings rather than represented as a QUBO. The resulting unconstrained objective is

$t$0

which preserves one qubit per asset even in the presence of multiple linear inequality constraints. This is a concrete departure from slack-variable-heavy QUBO reductions (Agliardi et al., 19 Aug 2025).

For continuous or oracle-based methods, the encoding is instead amplitude-based. Quantum linear-system approaches assume qRAM oracles for prices or returns, prepare a state of historical returns, and derive amplitude-encoded states for expected returns and covariance. The covariance matrix can appear as a density operator

$t$1

which is then processed with HHL-style or block-encoding/QSVT techniques to obtain a state proportional to the optimal portfolio (Rebentrost et al., 2018).

Online quantum portfolio optimization uses yet another encoding. It assumes online gain oracles

$t$2

constructs amplitude states approximately proportional to portfolio weights, estimates $t$3 norms and inner products via amplitude estimation, and samples assets from the distribution defined by the portfolio vector. The quantum state is therefore not only a representation of the solution but also the mechanism by which transaction-cost-reducing multi-sampling is implemented (Lim et al., 2022).

More recent scalable encodings target the variable-per-qubit bottleneck directly. Pauli Correlation Encoding maps many binary decisions to signs of expectation values of $t$4-local Pauli strings,

$t$5

with capacity

$t$6

For $t$7 and $t$8, the paper gives $t$9, enabling more than 250 variables on 9 qubits in principle (Soloviev et al., 26 Nov 2025).

3. Online, oracle-based, and theoretically guaranteed algorithms

The clearest regret-guarantee result in the literature provided here is the quantum version of the Exponentiated Gradient portfolio method. The classical baseline updates

$w^{(t)} \in \Delta^n$0

with initialization $w^{(t)} \in \Delta^n$1 and regret

$w^{(t)} \in \Delta^n$2

A sampling version invests only in $w^{(t)} \in \Delta^n$3 sampled assets per round, where

$w^{(t)} \in \Delta^n$4

reducing transaction cost to dependence on $w^{(t)} \in \Delta^n$5, $w^{(t)} \in \Delta^n$6, and $w^{(t)} \in \Delta^n$7, but not on $w^{(t)} \in \Delta^n$8 (Lim et al., 2022).

The quantum algorithm preserves that online setting and attains a quadratic speedup in the number of assets under the stated oracle model. Its main theorem gives

$w^{(t)} \in \Delta^n$9

with success probability at least $$S_{t+1} = S_t (w^{(t)} \cdot r^{(t)}), \qquad S_T = S_0 \prod_{t=1}^T (w^{(t)} \cdot r^{(t)}).$$0, and total runtime

$$S_{t+1} = S_t (w^{(t)} \cdot r^{(t)}), \qquad S_T = S_0 \prod_{t=1}^T (w^{(t)} \cdot r^{(t)}).$$1

The speedup comes from quantum norm estimation, relative-error inner-product estimation, and quantum multi-sampling, while the total transaction cost remains

$$S_{t+1} = S_t (w^{(t)} \cdot r^{(t)}), \qquad S_T = S_0 \prod_{t=1}^T (w^{(t)} \cdot r^{(t)}).$$2

independent of $$S_{t+1} = S_t (w^{(t)} \cdot r^{(t)}), \qquad S_T = S_0 \prod_{t=1}^T (w^{(t)} \cdot r^{(t)}).$$3 (Lim et al., 2022).

This line of work is tightly conditional on the data-access model. The same source states that constructing online gain oracles, for example via QRAM or coherent loading, is a major practical bottleneck, and that building QRAM for $$S_{t+1} = S_t (w^{(t)} \cdot r^{(t)}), \qquad S_T = S_0 \prod_{t=1}^T (w^{(t)} \cdot r^{(t)}).$$4 rounds with $$S_{t+1} = S_t (w^{(t)} \cdot r^{(t)}), \qquad S_T = S_0 \prod_{t=1}^T (w^{(t)} \cdot r^{(t)}).$$5 entries requires $$S_{t+1} = S_t (w^{(t)} \cdot r^{(t)}), \qquad S_T = S_0 \prod_{t=1}^T (w^{(t)} \cdot r^{(t)}).$$6 memory and preparation time $$S_{t+1} = S_t (w^{(t)} \cdot r^{(t)}), \qquad S_T = S_0 \prod_{t=1}^T (w^{(t)} \cdot r^{(t)}).$$7. It also states that the algorithm is best suited for fault-tolerant regimes because coherent amplitude estimation and repeated oracle use are difficult on NISQ devices (Lim et al., 2022).

The broader oracle-based mean-variance literature has a different theoretical emphasis. Quantum linear-system methods aim at preparing a state proportional to the optimal portfolio or at computing the efficient frontier with runtime polylogarithmic in the size of the historical return dataset after qRAM loading. Those methods rely on well-conditioned covariance structure, quantum access to returns or to covariance entries, and quantum state preparation for $$S_{t+1} = S_t (w^{(t)} \cdot r^{(t)}), \qquad S_T = S_0 \prod_{t=1}^T (w^{(t)} \cdot r^{(t)}).$$8 and $$S_{t+1} = S_t (w^{(t)} \cdot r^{(t)}), \qquad S_T = S_0 \prod_{t=1}^T (w^{(t)} \cdot r^{(t)}).$$9; they target global portfolio statistics and sampling access rather than direct enumeration of all weights (Rebentrost et al., 2018).

A plausible implication is that “quantum-enhanced portfolio optimization” contains two theoretically distinct notions of enhancement. One is query-speedup enhancement in oracle models with regret or runtime theorems; the other is search-heuristic enhancement in discrete, encoded optimization landscapes. The literature treats these as related but not interchangeable categories (Lim et al., 2022, Rebentrost et al., 2018).

4. Variational, annealing, walk-based, and counterdiabatic methods

For discrete selection, the dominant NISQ-era paradigm is variational optimization over Ising or HUBO encodings. Standard QAOA uses

$$L_T := \frac{1}{T} \sum_{t=1}^T \log(w^{(t)} \cdot r^{(t)}),$$0

with $$L_T := \frac{1}{T} \sum_{t=1}^T \log(w^{(t)} \cdot r^{(t)}),$$1 in the simplest case and

$$L_T := \frac{1}{T} \sum_{t=1}^T \log(w^{(t)} \cdot r^{(t)}),$$2

as the parameterized state. This is the template behind higher-moment HUBO QAOA, several mean-variance QUBO studies, and the expert-analysis study of QAOA versus SamplingVQE (Uotila et al., 1 Sep 2025, Innan et al., 28 Jul 2025).

One recurrent issue is that standard QAOA and VQE optimize the encoded cost, not necessarily a financially viable portfolio. In the expert-analysis evaluation study, both methods demonstrate effective cost-function minimization, but the resulting portfolios often violate practical finance criteria such as adequate diversification and realistic risk exposure. The paper therefore introduces an Expert Analysis Evaluation framework centered on liquidity, sector diversification, investor suitability, and short forward tests, and argues that expert judgment is necessary in the decision-making pipeline (Innan et al., 28 Jul 2025).

Several papers replace generic mixers by constraint-preserving dynamics. Quantum walk-based portfolio optimization uses ternary positions $$L_T := \frac{1}{T} \sum_{t=1}^T \log(w^{(t)} \cdot r^{(t)}),$$3, encodes long and short bits, and compares penalty-based QAOA, a feasibility-preserving QAOAz variant, and a Quantum Walk Optimization Algorithm defined on the feasible non-degenerate subspace. For $$L_T := \frac{1}{T} \sum_{t=1}^T \log(w^{(t)} \cdot r^{(t)}),$$4 and $$L_T := \frac{1}{T} \sum_{t=1}^T \log(w^{(t)} \cdot r^{(t)}),$$5, the paper reports search-space sizes of $$L_T := \frac{1}{T} \sum_{t=1}^T \log(w^{(t)} \cdot r^{(t)}),$$6 states for QAOA, $$L_T := \frac{1}{T} \sum_{t=1}^T \log(w^{(t)} \cdot r^{(t)}),$$7 valid encodings for QAOAz, and $$L_T := \frac{1}{T} \sum_{t=1}^T \log(w^{(t)} \cdot r^{(t)}),$$8 valid non-degenerate portfolios for QWOA, with QWOA showing the fastest convergence at low depth and the smallest variance across runs (Slate et al., 2020).

Counterdiabatic extensions pursue the same idea from an adiabatic-control perspective. Digitized-counterdiabatic portfolio optimization adds approximate counterdiabatic terms to a discretized adiabatic evolution, and the study reports drastic improvements in success probabilities relative to plain digitized evolution and improved performance relative to QAOA and DC-QAOA on the reported instances (Hegade et al., 2021). A newer constrained version, CCD-QAOA, builds approximate adiabatic gauge potentials from nested commutators of the Ising cost Hamiltonian and the Hamming-weight-preserving XY mixer. In simulations with $$L_T := \frac{1}{T} \sum_{t=1}^T \log(w^{(t)} \cdot r^{(t)}),$$9 assets and budget $w^T \Sigma w$0, it consistently yields better approximation ratios than standard XY-mixer QAOA, Grover-mixer QAOA, and penalty-based QAOA at fixed depth (Falla et al., 7 May 2026).

Another route is to modify the loss rather than the unitary family. In VQE for portfolio QUBO, the use of CVaR and especially Weighted CVaR replaces the plain expectation objective by a tail-focused objective

$w^T \Sigma w$1

and is paired with CMA-ES rather than COBYLA. In the reported 12-asset simulations, WCVaR combined with CMA-ES improves success rate and reduces sensitivity to the choice of $w^T \Sigma w$2 relative to unweighted CVaR (Wang et al., 26 Aug 2025).

5. Scalable encodings, hybrid workflows, and application-specific systems

A major practical theme is that realistic portfolio problems are too large or too constrained for naive one-variable-per-qubit encodings. One response is hybrid decomposition. The annealer–gate-model scheme built around Large System Sampling Approximation solves sub-QUBOs on a D-Wave annealer and recombines them through a gate-model VQE amplitude optimization. Its modification replaces random sampling by a Maximum Independent Set over a correlation-thresholded market graph, then builds representative sub-systems from highly correlated neighbors. On Indian stock data with up to 64 assets, the study reports approximation quality at par with classical methods while using substantially fewer sub-systems than random LSSA (Jain et al., 2023).

A related but distinct answer is automatic universe reduction before invoking the quantum solver. Q4FuturePOP first generates a predicted future dataset whose covariance matches the historical covariance while cumulative expected returns match expert-provided future expectations, then repeatedly solves the portfolio problem on the full universe to identify assets that consistently receive nonzero allocations, and finally solves the reduced problem on a D-Wave Advantage annealer. This is an explicitly forward-looking system rather than a purely historical-return optimizer (Osaba et al., 2023).

Scalability can also be attacked through representation. Pauli Correlation Encoding compresses a graph partitioning problem over pairwise asset correlations into expectation values of multi-qubit Pauli strings, then recursively bipartitions the market graph and selects one representative asset per cluster by expected return. In the reported S&P 500 experiments, PCE handles graph sizes up to 250 variables with gate counts under about 750 for $w^T \Sigma w$3, whereas QAOA baselines at similar graph density are much larger; the portfolio evaluated by equally weighting cluster representatives outperforms the baseline “full market” allocation and exhibits superior Sharpe ratios on train and test sets (Soloviev et al., 26 Nov 2025).

For realistic ETF bond portfolio design, the most application-specific workflow is the sampling-based CVaR-VQA combined with local search. The formulation targets class-level metric matching with multiple guardrails, evaluates the cost directly on samples, and demonstrates utility-scale experiments on IBM Heron processors with 109 qubits and up to about 4,200 gates. The best reported hardware relative solution error after local-search post-processing is $w^T \Sigma w$4, and the combined quantum-classical workflow outperforms purely classical local search on the reported 100+ variable instances (Agliardi et al., 19 Aug 2025).

This suggests that scalability in the current literature is obtained less by a single universal quantum primitive than by problem restructuring: reduction of the investment universe, hierarchical decomposition, compressed encodings, or classical post-processing layered on top of variational sampling. The common pattern is hybridization rather than standalone quantum execution (Jain et al., 2023, Osaba et al., 2023, Soloviev et al., 26 Nov 2025, Agliardi et al., 19 Aug 2025).

6. Empirical status, practical bottlenecks, and open issues

The empirical record is mixed and method-dependent. Some results are primarily theoretical, such as the online regret theorems for the quantum Exponentiated Gradient variant (Lim et al., 2022). Some are simulator-heavy, such as higher-moment HUBO-QAOA, WCVaR-VQE, and many QAOA/VQE circuit studies (Uotila et al., 1 Sep 2025, Wang et al., 26 Aug 2025, Zaman et al., 2024). Others report small or medium-size hardware demonstrations. Black-Litterman with QML views reports a 12-qubit, choose-6 example and a real-device test of final VQE sampling on IBM Auckland; the mean certainty-equivalent return for the 12-asset case is reported as $w^T \Sigma w$5 for BL VQE versus $w^T \Sigma w$6 for exact BL and $w^T \Sigma w$7 for naive $w^T \Sigma w$8 (Chen et al., 2023). The ETF bond study reports 109-qubit hardware runs on IBM Heron processors and a $w^T \Sigma w$9 relative solution error after local search (Agliardi et al., 19 Aug 2025).

Annealing-based realistic studies also exist. In an end-to-end Indian equity workflow, the algorithm’s allocation is reported to achieve returns of $\mu^T w = \mu_p$0 versus $\mu^T w = \mu_p$1 for the benchmark, risk of $\mu^T w = \mu_p$2 versus $\mu^T w = \mu_p$3, Sharpe ratio of $\mu^T w = \mu_p$4 versus $\mu^T w = \mu_p$5, and diversification ratio of $\mu^T w = \mu_p$6 versus $\mu^T w = \mu_p$7; the abstract summarizes a portfolio increase of 200,000 INR over the benchmark, while the figure indicates an improvement of about 0.25–0.30 million INR over 13 months (Morapakula et al., 10 Apr 2025). ESG-constrained discrete Markowitz on EURO STOXX 50 subsets reports that directly optimized discrete portfolios lie on the efficient frontier and can outperform simple rounding of the continuous solution for moderate budgets, while progressively tighter ESG constraints move solutions away from the unconstrained frontier (Catalano et al., 2024).

At the same time, the literature repeatedly stresses bottlenecks and negative findings. The higher-moment HUBO paper states that exact eigensolvers often demonstrate better allocations than the classical baseline, but QAOA, as deployed, matches exact or classical quality only on a subset of instances, especially as qubit count grows (Uotila et al., 1 Sep 2025). The expert-analysis study finds a critical disparity between algorithmic performance and financial applicability (Innan et al., 28 Jul 2025). The constrained DGMVP QAOA study states that realistic thermal relaxation noise levels preclude quantum advantage, although in a regime where hardware improves enough for stochastic measurement noise to dominate, the required number of shots to obtain the global minimum scales favorably in the reported simulations (Yuan et al., 2024).

A frequent misconception is that better encoded energy automatically means better portfolio construction. The studies provided here consistently reject that view. Financially meaningful deployment depends on data-access assumptions, feasibility enforcement, noise tolerance, budget and transaction models, diversification structure, and in several cases explicit expert or post-processing filters (Lim et al., 2022, Innan et al., 28 Jul 2025, Agliardi et al., 19 Aug 2025). Another misconception is that all quantum portfolio papers address the same financial problem. The corpus instead covers adversarial online wealth maximization, batch mean-variance optimization, higher-order moment selection, ESG-constrained discrete allocation, correlation-graph partitioning, and Black-Litterman inference, with different objectives, constraints, and output semantics (Rebentrost et al., 2018, Uotila et al., 1 Sep 2025, Catalano et al., 2024, Soloviev et al., 26 Nov 2025, Chen et al., 2023).

Open problems are stated explicitly across the sources. They include reducing the $\mu^T w = \mu_p$8 dependence in online oracle-based methods, practical QRAM-free implementations, stronger comparative numerical evaluation against classical baselines, better constraint-preserving mixers and warm starts, robust estimation of skewness and kurtosis, integration of transaction costs and turnover, adaptive penalty calibration, larger-scale hardware demonstrations, and hybrid workflows that combine quantum subroutines with classical convex or local-search refinements (Lim et al., 2022, Uotila et al., 1 Sep 2025, Agliardi et al., 19 Aug 2025, Falla et al., 7 May 2026). The present state of the field therefore supports a precise characterization: quantum-enhanced portfolio optimization is an active set of algorithmic and modeling strategies with mathematically diverse formulations and some promising empirical results, but with practical advantage still strongly conditional on the financial model, the encoding, the data-access regime, and the hardware assumptions.