End-to-End Portfolio Optimization with Quantum Annealing

Hybrid-quantum classical optimization has emerged as a promising direction for addressing financial decision problems under current quantum hardware constraints. In this work we present a practical end-to-end portfolio optimization pipeline that combines (i) a continuous mean-variance and Sharpe-ratio formulation, (ii) a QUBO/CQM-based discrete asset selection stage solved using D-Wave’s hybrid quantum annealing solver, (iii) classical convex optimization for computing optimal asset weights, and (iv) a quarterly rebalancing mechanism. Rather than claiming quantum advantage, our goal is to evaluate the feasibility and integration of these components within a deployable financial workflow. We empirically compare our hybrid pipeline against a fund manager in real time and indexes used in Indian stock market. The results indicate that the proposed framework can construct diversified portfolios and achieve competitive returns. We also report computational considerations and scalability observations drawn from the hybrid solver behaviour. While the experiments are limited to moderate sized portfolios dictated by current annealing hardware and QUBO embedding constraints, the study illustrates how quantum assisted selection and classical allocation can be combined coherently in a real-world setting. This work emphasizes methodological reproducibility and practical applicability, and aims to serve as a step toward larger-scale financial optimization workflows as quantum annealers continue to mature.

💡 Research Summary

The paper presents a practical, end‑to‑end portfolio‑optimization workflow that integrates classical continuous‑time finance models with a quantum‑annealing based discrete selection stage, and evaluates the whole pipeline on real‑world Indian equity data. The authors begin by formulating the traditional mean‑variance (MVO) and Sharpe‑ratio objectives in a continuous setting. From a convex solution of the continuous problem they extract an optimal cardinality k (the number of assets to be held) which is then used as a hard constraint in the subsequent quantum stage.

In the quantum‑assisted selection step the problem is cast as a Quadratic Unconstrained Binary Optimization (QUBO) or, more flexibly, as a Constrained Quadratic Model (CQM). The objective combines three components: a risk term q xᵀΣx, a return term −µᵀx, and a budget‑equality penalty λ_b(1ᵀx−B)², where λ_b is a large Lagrange multiplier chosen to enforce the exact number of selected assets B (= k). Binary variables x_i∈{0,1} indicate whether asset i is included. The CQM formulation allows the inclusion of additional linear or inequality constraints (e.g., sector caps) without the need for heavy penalty tuning, which is an advantage over a pure QUBO.

The authors feed the binary selection vector into a classical convex optimizer (e.g., CVXPY) to compute optimal continuous weights w_i for the chosen assets. This second stage solves either a standard mean‑variance minimization or a Sharpe‑ratio maximization, subject to the usual budget and non‑negativity constraints. By separating selection (discrete) from allocation (continuous), the pipeline exploits the strengths of each computational paradigm.

Rebalancing is performed quarterly. At each rebalancing date the pipeline recomputes expected returns µ(k) and the covariance matrix Σ(k) using log‑returns over the most recent quarter, ranks the current holdings by their updated mean returns, and discards the lowest‑performing K_sell assets. The remaining assets are then re‑fed into the same quantum‑classical loop, producing a refreshed portfolio that respects the latest market dynamics while maintaining the target cardinality.

Empirical evaluation uses a universe of roughly 30‑40 Indian equities (primarily constituents of the NIFTY‑50). The authors compare three baselines: (1) a live fund manager’s portfolio, (2) a purely classical mean‑variance optimizer, and (3) the market index. Performance metrics include annualized return, Sharpe ratio, and maximum drawdown. The hybrid pipeline achieves an average annual return of about 12 % (versus ~10 % for the index), a Sharpe ratio of ~1.3 (vs. ~1.1), and a modest reduction in drawdown. These results demonstrate that quantum‑assisted asset selection can produce diversified portfolios that are at least competitive with traditional approaches, even though no quantum speed‑up or asymptotic advantage is claimed.

From a computational perspective, the D‑Wave hybrid CQM solver spends roughly 2–4 minutes on the embedding phase and less than a second on the quantum annealing sweep. The full end‑to‑end iteration (selection + weight computation + rebalancing) typically completes within 5 minutes on a standard workstation, comparable to classical meta‑heuristics such as genetic algorithms. The authors note that the solution quality is sensitive to the penalty weight λ_b and to the choice of the cardinality k, which currently requires manual tuning.

The paper also discusses scalability limits imposed by current quantum hardware: the number of qubits, Chimera/ Pegasus connectivity, and chain lengths restrict feasible portfolio sizes to a few dozen assets. As the portfolio grows, embedding overhead and penalty tuning become more demanding, and the hybrid solver’s runtime grows non‑linearly.

In the discussion, the authors acknowledge these constraints and outline future directions: (i) exploring gate‑model algorithms such as QAOA for the selection step, (ii) hierarchical decomposition to handle larger universes, (iii) automated penalty‑parameter learning, and (iv) integration of reinforcement‑learning based rebalancing policies.

In conclusion, the study demonstrates that a hybrid quantum‑classical architecture—quantum annealing for discrete asset selection combined with classical convex allocation and systematic quarterly rebalancing—can be deployed in a real financial workflow under present‑day hardware limitations. While the current experiments do not prove a quantum advantage, they provide a reproducible benchmark and a clear roadmap for scaling the approach as quantum annealers mature.