This thesis studies the formulation cost of applying binary optimization to dynamic asset allocation as preparation for quantum and quantum-inspired solution methods. The financial benchmark is a monthly rolling-horizon mean-variance model with a three-month planning horizon, proportional transaction costs, full investment, long-only and upper-bound restrictions, and a turnover limit. Expected returns and covariances are estimated from rolling 36-month windows for an eight-ETF universe. The analysis proceeds from the continuous model to a global trade grid and then to a local Quadratic Unconstrained Binary Optimization (QUBO) model centered on the continuous trade path. Refining the coarse global grids improves the sampled instance where the returned path is poorly aligned with the continuous benchmark, but the finest tested grid adds runtime without further improvement. Because the grid problem is solved heuristically, its observed gaps combine representation and search error. Constraint ablation then shows the separate cost of QUBO translation: 24 economic trade decisions become 423 binary variables and 5,227 nonzero quadratic terms in the complete local instance. Simulated annealing and greedy coordinate search are used as transparent classical diagnostics; no quantum advantage is claimed. Across a penalty sweep and nine sampled rebalance dates, centered runs preserve decoded feasibility while random valid one-hot starts remain unreliable. The center is included in the local discrete domain, so a feasible local solution is not expected to improve a globally optimal continuous benchmark under the same objective. The defensible role of a future quantum backend is therefore to solve a clearly specified binary subproblem and to be compared with classical methods on that same formulation.
This thesis studies the formulation cost of applying binary optimization to dynamic asset allocation as preparation for quantum and quantum-inspired solution methods. The financial benchmark is a monthly rolling-horizon mean-variance model with a three-month planning horizon, proportional transaction costs, full investment, long-only and upper-bound restrictions, and a turnover limit. Expected returns and covariances are estimated from rolling 36-month windows for an eight-ETF universe. The analysis proceeds from the continuous model to a global trade grid and then to a local Quadratic Unconstrained Binary Optimization (QUBO) model centered on the continuous trade path. Refining the coarse global grids improves the sampled instance where the returned path is poorly aligned with the continuous benchmark, but the finest tested grid adds runtime without further improvement. Because the grid problem is solved heuristically, its observed gaps combine representation and search error. Constraint ablation then shows the separate cost of QUBO translation: 24 economic trade decisions become 423 binary variables and 5,227 nonzero quadratic terms in the complete local instance. Simulated annealing and greedy coordinate search are used as transparent classical diagnostics; no quantum advantage is claimed. Across a penalty sweep and nine sampled rebalance dates, centered runs preserve decoded feasibility while random valid one-hot starts remain unreliable. The center is included in the local discrete domain, so a feasible local solution is not expected to improve a globally optimal continuous benchmark under the same objective. The defensible role of a future quantum backend is therefore to solve a clearly specified binary subproblem and to be compared with classical methods on that same formulation.
Dynamic Asset Allocation with Quantum Optimization
CREMA, ALESSANDRO
2025/2026
Abstract
This thesis studies the formulation cost of applying binary optimization to dynamic asset allocation as preparation for quantum and quantum-inspired solution methods. The financial benchmark is a monthly rolling-horizon mean-variance model with a three-month planning horizon, proportional transaction costs, full investment, long-only and upper-bound restrictions, and a turnover limit. Expected returns and covariances are estimated from rolling 36-month windows for an eight-ETF universe. The analysis proceeds from the continuous model to a global trade grid and then to a local Quadratic Unconstrained Binary Optimization (QUBO) model centered on the continuous trade path. Refining the coarse global grids improves the sampled instance where the returned path is poorly aligned with the continuous benchmark, but the finest tested grid adds runtime without further improvement. Because the grid problem is solved heuristically, its observed gaps combine representation and search error. Constraint ablation then shows the separate cost of QUBO translation: 24 economic trade decisions become 423 binary variables and 5,227 nonzero quadratic terms in the complete local instance. Simulated annealing and greedy coordinate search are used as transparent classical diagnostics; no quantum advantage is claimed. Across a penalty sweep and nine sampled rebalance dates, centered runs preserve decoded feasibility while random valid one-hot starts remain unreliable. The center is included in the local discrete domain, so a feasible local solution is not expected to improve a globally optimal continuous benchmark under the same objective. The defensible role of a future quantum backend is therefore to solve a clearly specified binary subproblem and to be compared with classical methods on that same formulation.| File | Dimensione | Formato | |
|---|---|---|---|
|
Crema_Alessandro.pdf
Accesso riservato
Dimensione
2.46 MB
Formato
Adobe PDF
|
2.46 MB | Adobe PDF |
The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License
https://hdl.handle.net/20.500.12608/115043