quantum-portfolio-optimization

# Quantum Portfolio Optimization Quantum computing methodologies for portfolio optimization — covering QAOA formulations with higher-order moments and quantum annealing pipelines for NISQ-era financial decision making. ## Core Papers ### QAOA for Higher-Order Portfolio Optimization (arXiv: 2509.01496) First quantum formulation for portfolio optimization with **higher-order moments** (skewness and kurtosis). Standard mean-variance ignores distribution asymmetry and tail risk. QAOA encodes cubic/quadratic terms into Ising Hamiltonians, enabling quantum advantage for complex risk modeling. **Key Insight**: Including skewness (3rd moment) and kurtosis (4th moment) in the objective function leads to better risk-adjusted returns. The QAOA circuit depth scales polynomially with the number of assets when using higher-order terms. ### End-to-End Quantum Annealing Pipeline (arXiv: 2504.08843) Practical hybrid pipeline combining: 1. Continuous mean-variance and Sharpe-ratio objectives (classical preprocessing) 2. QUBO formulation for quantum annealing solver 3. Post-processing and validation on classical hardware Demonstrates feasibility on current NISQ devices. ## Usage Patterns ### Pattern 1: QAOA Higher-Order Portfolio Optimization Use QAOA when: - Portfolio needs risk modeling beyond mean-variance - Skewness/kurtosis matter for the asset class - Quantum hardware access available (or simulator for small problems) **QUBO formulation:** ``` H = -μ^T x + λ₁ x^T Σ x + λ₂ Σᵢⱼₖ Sᵢⱼₖ xᵢxⱼxₖ + λ₃ Σᵢⱼₖₗ Kᵢⱼₖₗ xᵢxⱼxₖxₗ ``` Where S = skewness tensor, K = kurtosis tensor, x = binary selection vector. **QAOA steps:** 1. Encode objective as Ising Hamiltonian 2. Map to QUBO with penalty terms for constraints 3. Initialize QAOA with p layers (p=1-3 for NISQ) 4. Optimize angles classically (COBYLA/SPSA) 5. Sample from final state for portfolio candidates ### Quantum Annealing Pipeline (UPDATED 2026-05-30) ⚠️ **Important caveat**: See Error Handling section for critical findings on penalty-encoded QUBO failure and D-Wave audit. Use constraint-native CQM instead of penalty QUBO. Use quantum annealing when: - Problem size fits current QA hardware (D-Wave: ~5000 qubits) - Need practical results on existing hardware - Mean-variance + Sharpe ratio sufficient (no higher moments) **Pipeline (revised):** 1. **Classical preprocessing**: Compute returns, covariance, constraints 2. **Constraint-native CQM formulation**: Use D-Wave's CQM interface, NOT penalty-encoded QUBO 3. **Embedding**: Let CQM solver handle constraints natively 4. **Annealing**: Run multiple reads (1000-10000 samples) 5. **Post-processing**: Select best feasible solution, validate Sharpe ratio 6. **Audit**: Report actual QPU time fraction vs wall-clock time (expect <1%) ### Pattern 3: Hybrid Classical-Quantum For production-grade portfolio optimization: 1. Use classical optimization for initial solution 2. Use quantum (QAOA/QA) for refinement in local neighborhoods 3. Validate against classical benchmarks 4. Track quantum advantage as hardware improves ## Mathematical Framework ### QUBO Encoding ```python import numpy as np from typing import Tuple def portfolio_to_qubo( returns: np.ndarray, covariance: np.ndarray, risk_aversion: float = 1.0, budget: int = None, penalty: float = 10.0 ) -> Tuple[np.ndarray, float]: """Encode portfolio optimization as QUBO matrix. H = -μ'x + λ·x'Σx + P·(Σxᵢ - K)² """ n = len(returns) if budget is None: budget = n // 2 # Objective: -μ'x + λ·x'Σx Q = risk_aversion * covariance - np.outer(returns, np.ones(n)) * 0.5 Q = Q + Q.T # symmetrize # Budget constraint: (Σxᵢ - K)² Q += penalty * np.ones((n, n)) Q -= penalty * budget * np.eye(n) offset = penalty * budget**2 return Q, offset ``` ### QAOA Circuit Construction ```python # Pseudocode for QAOA portfolio circuit # 1. Initialize: |+⟩^⊗n (equal superposition over all portfolios) # 2. For p layers: # a. Apply cost Hamiltonian: exp(-i·γ·H_C) # - H_C = Σᵢ (-μᵢ + λ·σᵢᵢ) Zᵢ + Σᵢ<ⱼ λ·σᵢⱼ ZᵢZⱼ + ... # b. Apply mixer: exp(-i·β·Σᵢ Xᵢ) # 3. Measure in computational basis # 4. Classical optimizer updates (γ, β) ``` ## Error Handling ### ⚠️ CRITICAL: Penalty-Encoded QUBO Failure on D-Wave (arXiv: 2605.17628, 2605.17623) - **Problem**: Standard penalty-encoded QUBO portfolio optimization **fails structurally** on current D-Wave Pegasus/Zephyr hardware - **Root cause**: Cardinality penalty contributes dense rank-one term (proportional to all-ones matrix), making logical interaction graph complete regardless of covariance structure - **Symptom**: Chain-break fractions reach 83% at N=24 and 92% at N=48 - **Fix**: Reformulate as **constraint-native CQM** (Constrained Quadratic Model) instead of penalty-encoded QUBO - **Audit finding**: D-Wave LeapHybridCQM matches Gurobi optimum but QPU access is only 0.034s out of 5s budget (0.7%) — quantum contribution is marginal at current scale - **Recommendation**: Use classical MIQP (Gurobi/CPLEX) for production; reserve quantum for research/hardware evolution tracking ### QAOA Barren Plateaus - **Symptom**: Cost function gradients vanish exponentially with qubit count - **Mitigation**: Use problem-specific initialization, layerwise training (p=1 → p=2 → ...) - **Reference**: See `quantum-neural-barren-plateau` skill ### Quantum Annealing Embedding Failures - **Symptom**: QUBO doesn't fit hardware topology - **Mitigation**: Use chain strength optimization, problem decomposition, or classical post-processing ### Higher-Order Moment Estimation - **Symptom**: Noisy skewness/kurtosis estimates from limited data - **Mitigation**: Use shrinkage estimators, Bayesian priors, or robust statistics ## Activation Keywords - quantum portfolio optimization - QAOA finance - quantum annealing portfolio - higher-order moment portfolio - quantum finance optimization - 量子组合优化 - QAOA 投资组合 - 量子退火金融 - 量子金融优化 ## Related Skills - `quantum-optimization-qaoa` — General QAOA methodology - `quantum-finance-portfolio` — Broader quantum finance patterns - `quantum-neural-barren-plateau` — Barren plateau mitigation