Toward Quantum Utility in Finance: A Robust Data-Driven Algorithm for Asset Clustering

1. Problem Statement

The paper addresses the challenge of clustering financial assets based on return correlations to optimize portfolio construction and statistical arbitrage.

• The Core Difficulty: Financial assets exhibit complex signed correlation structures (positive correlations indicating co-movement and negative correlations indicating anti-correlation).
• Limitations of Classical Methods:
  • Lossy Transformations: Traditional algorithms (e.g., k-Means, k-Medoids, hierarchical clustering) require converting signed correlations ($\rho \in [-1, 1]$) into non-negative distances. This transformation (e.g., $d_{ij} = \sqrt{\alpha(1-\rho_{ij})}$) destroys semantic fidelity; for instance, a zero correlation is mapped to a fixed non-zero distance, misrepresenting the data.
  • Centroid Deviation: Centroid-based methods minimize distance to a center rather than maximizing intra-cluster correlations and minimizing inter-cluster correlations.
  • Hyperparameter Sensitivity: Classical methods require manual tuning of the number of clusters ($k$), initialization strategies, or merging thresholds, which are not generalizable across different datasets.
  • Spectral Limitations: While spectral methods like SPONGE handle signed graphs, they perform best on sparse graphs with discrete weights, conditions rarely met in dense financial data.
• Quantum Context: Previous quantum attempts often formulated clustering as a Maximum Clique problem, requiring excessive binary variables that exceed current quantum hardware capabilities.

2. Methodology: GCS-Q Algorithm

The authors propose GCS-Q (Graph-based Coalition Structure Generation), a hybrid quantum-classical algorithm designed to operate directly on signed, weighted graphs without lossy transformations.

• Graph Formulation:

  • Assets are nodes ($V$); pairwise Pearson correlations are edge weights ($w_{ij} = \rho_{ij} \in [-1, 1]$).
  • Objective: Find a partition $\Pi = \{C_1, \dots, C_k\}$ that maximizes the sum of intra-cluster edge weights:
    $$\max_{\Pi} \sum_{C \in \Pi} \sum_{i,j \in C, i<j} w_{ij}$$
  • This naturally groups positively correlated assets and separates negatively correlated ones.

• Iterative Process:

  1. Initialization: Start with the full graph $G_0 = G$.
  2. Minimum Cut: Iteratively compute a minimum cut $(S_t, \bar{S}_t)$ on the current subgraph. This step is NP-Hard.
  3. Recursive Partitioning: The graph is split into subgraphs induced by $S_t$ and $\bar{S}_t$. Cut edges are removed.
  4. Stopping Criterion: The recursion stops when no cut exists with a value lower than the sum of all edges in the subgraph. This dynamically determines $k$, eliminating the need for manual tuning.

• Quantum Implementation:

  • The minimum cut subproblem is formulated as a Quadratic Unconstrained Binary Optimization (QUBO) problem.
  • It is solved using D-Wave's quantum annealer (Advantage_system5.4), which explores the exponentially large solution space more efficiently than classical solvers for this specific combinatorial problem.
  • The workflow is hybrid: the iterative logic runs on a classical CPU, while the QUBO solving is offloaded to the quantum annealer.

3. Key Contributions

1. Direct Signed Graph Processing: GCS-Q is the first to apply a quantum algorithm to asset clustering without transforming signed correlations into non-negative distances, preserving the true relational structure of financial data.
2. Dynamic Cluster Determination: The algorithm automatically infers the optimal number of clusters ($k$) based on the graph's structural balance, removing the need for heuristic parameter tuning (e.g., elbow method).
3. Demonstration of Quantum Utility: The paper provides empirical evidence of quantum advantage in solution quality (not just speed) using real-world quantum hardware (D-Wave) for a practical financial application.
4. Robustness to Real-World Data: Unlike prior spectral methods requiring sparse or discrete graphs, GCS-Q handles dense, continuous-weight graphs with heterogeneous cluster sizes.

4. Experimental Results

The authors benchmarked GCS-Q against SPONGE (and its symmetric variant) and k-Medoids (PAM) using two datasets:

• Synthetic Data:

  • Setup: Structurally balanced signed graphs with continuous weights and non-uniform cluster sizes (mimicking financial markets).
  • Metric: Adjusted Rand Index (ARI).
  • Outcome: GCS-Q consistently achieved higher ARI scores than classical baselines. PAM performed poorly due to distance transformation issues, while SPONGE degraded due to its reliance on sparsity and discrete weights.

• Real-World Data (Yahoo Finance):

  • Setup: Hourly returns of 50 assets across diverse sectors (January 2025).
  • Metric: Penalty Metric, which penalizes negative intra-cluster correlations and positive inter-cluster correlations (lower is better).
  • Outcome: GCS-Q achieved the lowest penalty, indicating superior structural balance. This suggests the clusters are internally cohesive (good for mean-reversion) and externally distinct (good for diversification).

• Performance Note: While the quantum annealer solves individual QUBOs quickly (capped at 1s), the total runtime for $N=170$ assets was ~10 minutes due to cloud latency and queue times. However, the focus was on solution quality, where GCS-Q outperformed classical methods.

5. Significance and Future Directions

• Practical Quantum Utility: This work marks a concrete step toward "Quantum Utility" in finance, demonstrating that near-term quantum annealers can solve specific, high-value unsupervised learning tasks better than classical heuristics.
• Portfolio Optimization: By producing structurally balanced clusters, GCS-Q enables more effective asset diversification and risk management, directly supporting the Markowitz mean-variance framework.
• Future Work:
  • Applying clusters to downstream tasks like asset allocation and evaluating performance via Sharpe Ratio.
  • Leveraging next-generation hardware (e.g., D-Wave's Zephyr topology) for better connectivity.
  • Exploring gate-model quantum solvers (QAOA) and hybrid embeddings.
  • Benchmarking against classical combinatorial optimizers (Gurobi, CPLEX).

In conclusion, the paper establishes that GCS-Q is a robust, data-driven algorithm that overcomes the limitations of classical clustering on signed graphs, offering a tangible quantum advantage in financial asset clustering.