Portfolio optimization is one of the quietest jobs in a bank and one of the most consequential. The textbook version — expected returns, a covariance matrix, a risk budget — is solved in milliseconds for thousands of assets. So why are quant research teams paying for time on quantum processors? Because the problems that actually reach a portfolio manager are not the textbook version. Add a cardinality cap, minimum lot sizes, turnover penalties, multi-period rebalancing and a drawdown constraint, and an easy convex problem turns combinatorial. Quantum hardware is one candidate for those cases. The pilots so far are more instructive than triumphant.

What banks are actually testing

Most publicly described banking work sits in four buckets:

• QUBO and Ising formulations of asset selection, run on quantum annealers
• Variational algorithms such as QAOA on gate-based hardware
• Quantum Monte Carlo for risk measures and derivatives pricing
• Hybrid loops where a quantum device proposes candidates and a classical solver refines them

JPMorgan Chase runs one of the larger in-house quantum research groups in the industry and has published on quantum optimization and Monte Carlo methods. Goldman Sachs worked with QC Ware and AWS on quantum approaches to Monte Carlo simulation. Several other global banks have announced vendor partnerships or joined hardware networks, often with IBM, D-Wave, IonQ, Quantinuum, QC Ware or Multiverse Computing.

Almost none of this is production. The instances are small — tens of variables, occasionally a few hundred — and many results come from simulators rather than real devices, because simulators let you isolate algorithmic behavior from hardware noise.

The data constraints nobody puts in the slide deck

Getting a bank's real portfolio onto a quantum processor is mostly a governance exercise, not a physics one.

• Confidentiality. Position and client data are sensitive, sometimes inside information. Sending a live book to a cloud quantum service triggers review by legal, compliance and model risk.
• Licensing. Market data carries redistribution limits, so a vendor may not legally receive the exact dataset a pilot needs. Teams substitute synthetic series or public proxies, which changes the problem.
• Model governance. Anything that influences a trade needs documentation, validation and an audit trail. Research notebooks rarely clear that bar.
• Metered access. Queue times and per-shot pricing shape experiment design more than most papers admit.

There is a technical constraint too. Annealers need extra physical qubits to represent connections between variables, so a 60-variable selection problem can consume far more hardware than expected, and chain breaks introduce errors. That forces coarse discretization of portfolio weights — which can change the optimal answer before quantum hardware is even involved.

Why classical methods are still hard to beat

Plain mean-variance optimization is a convex quadratic program. Interior-point and active-set solvers handle it reliably, and the answer is optimal to a stated tolerance. No quantum device has an edge there, and it is not obvious one will appear.

The constrained versions are harder, but they are not defenseless. Mixed-integer solvers such as Gurobi and CPLEX, plus decades of tuned heuristics, have absorbed enormous engineering effort. Simulated annealing, tabu search and large-neighborhood search attack the same combinatorial structure. Quantum annealers are physically annealers, so the fair benchmark is a well-tuned classical annealer on the identical QUBO. At sizes that fit today's hardware, the classical run usually matches or wins.

Three practical points that skew comparisons

Noise and calibration drift mean repeated runs return different answers, so you need many shots and honest statistics. Wall-clock time includes queueing and compilation, not just sampling. And a classical solver running on existing infrastructure costs nearly nothing per run, while cloud quantum access has a price per shot. Cost per unit of solution improvement is a better yardstick than any headline speedup.

Where a quantum approach could plausibly earn its place

The interesting cases are the ones where the problem stops being convex:

• Cardinality-constrained selection — hold exactly K of N assets
• Integer lot sizes and minimum trade sizes
• Turnover penalties and market-impact costs spread across many names
• Multi-period rebalancing with path dependence, tax lots and loss carryforward
• CVaR, drawdown or tracking-error constraints layered onto a budget
• Scenario-based robust allocation with a large scenario set

Even here, the honest framing is "a new heuristic candidate," not a breakthrough. Some teams get real value from quantum-inspired classical methods — tensor networks, specialized samplers, FPGA annealers — and never touch a QPU. That is a legitimate outcome, not a failure.

What a credible pilot looks like

1. Pick one constrained subproblem rather than the whole book. Asset selection with a cardinality cap is the usual starting point.
2. Formulate the QUBO, then solve the same instance with a strong classical baseline: a MIP solver under a time limit, plus a tuned heuristic.
3. Compare on identical instances and identical time budgets, counting queue and compilation time on the quantum side.
4. Run enough shots and seeds to report a distribution. A single good sample proves nothing.
5. Repeat with and without error mitigation, and check how behavior changes as the instance grows.
6. Judge by cost per unit of quality. If the quantum route needs ten times the budget for the same answer, that is the finding.

Warm-starting is the most plausible near-term production pattern: let quantum sampling generate candidate portfolios, then have a classical solver refine and verify them. The quantum device becomes a generator, not an oracle.

What to do with all this

If you work on a portfolio or risk team, build the classical baseline first and build it properly — an exact solver plus a tuned heuristic, with clear timing. Keep a short, documented library of the constrained problems your desk actually struggles with; that library is the asset that matters, whichever hardware eventually wins. Clean formulations travel well, and they let you evaluate any new solver, quantum or otherwise, in days instead of quarters.

Watch for error-corrected machines that can hold a few thousand logical qubits and run deeper circuits than today's devices. Until that happens, treat vendor announcements and preprint results as research progress, not as investment guidance, and keep expectations calibrated against the solvers already sitting on your cluster.

Photo: Masood Aslami / Pexels