What you'll take away
- There is no single best method — there is a best method for a given problem shape. The art is matching the family to the constraints you actually have.
- Exactness and scale are usually in tension. The methods that guarantee the true optimum are the ones that struggle to run across a large book; the methods that scale give up the guarantee.
- Taxes and discrete rules are what make real portfolios hard. They are exactly the features that break the cleanest, fastest methods.
Portfolio optimization sounds like one problem. In practice it is a family of related problems that differ along a few axes: whether the objective is smooth, whether the variables are continuous or have to be whole, how many constraints there are and how nasty they are, and — the axis that quietly dominates everything in wealth management — whether taxes are in the picture. The choice of method is really a choice about which of these features you are willing to model faithfully and which you are willing to approximate. Every honest practitioner is making that trade; the dishonest ones just are not telling you.
Family one: convex methods
The convex family is where most people start, and for good reason. When the objective is a smooth trade-off between expected return and risk, and the constraints are linear or otherwise well-behaved, the resulting problem has a single global optimum and a rich theory for finding it efficiently. This is the world of mean-variance optimization in its classical form, the direct descendant of Markowitz's 1952 framework. Convex problems are wonderful: they are fast, they are reliable, and when a convex solver returns an answer you can trust that it is the answer, not just an answer.
The catch is that the real world keeps refusing to be convex. The moment you need to say "hold no more than forty names," or "either own at least one percent of this position or none of it," or "harvest this loss but respect the wash-sale window," you have left the convex world, because those are discrete, combinatorial conditions and convexity cannot represent them. You can sometimes approximate your way around the edges, but the approximation is exactly where the value in tax-aware investing lives, which is why convex-only approaches systematically leave after-tax money on the table.
Family two: integer and combinatorial methods
When the problem genuinely has whole-number or either/or structure — position-count limits, minimum position sizes, discrete lot decisions — you are in the territory of integer programming and its relatives. These methods can model the real constraints faithfully and, given enough time, return a provably optimal answer. For a single account where exactness is worth waiting for, this is the gold standard, and nothing else beats it on its home turf.
The price is computational. Integer problems are, in the worst case, exponentially hard, and that worst case shows up exactly when the problems get large or the constraints get tight. A method that returns the exact answer to a small problem in a second can stall for minutes — or fail to finish at all — on a problem an order of magnitude bigger. This is not a flaw in any particular implementation; it is the nature of the problem class. The practical consequence is blunt: exact combinatorial methods do not run across a book of a hundred thousand accounts before the open, and no amount of engineering changes the underlying complexity.
Figure 1 — The exactness / scale trade-off, conceptually
A conceptual map, not a measurement. Up-and-to-the-right is the dream and the open research frontier: faithful answers and book-level scale at once.
Family three: first-order methods
Between the exact-but-slow and the fast-but-approximate poles sits a family of first-order methods that trade a little precision for a lot of speed and, crucially, for the ability to scale to very large problems and to run well on parallel hardware. These methods iterate toward a solution using gradient information rather than the heavier machinery exact solvers rely on, which makes each step cheap and the whole process amenable to the kind of massively parallel computation a modern accelerator is built for. For large continuous problems where a near-optimal answer delivered quickly beats an exact answer delivered late, this is often the sweet spot.
The honest caveat is that "near-optimal" has to be earned and measured, not assumed. A first-order method that is sloppy about its stopping criteria can return something that looks like an answer but violates a constraint or sits far from the true optimum. The discipline that separates a credible first-order approach from a dangerous one is rigorous quality checking against a known-good reference — which is a recurring theme in any serious benchmark methodology.
Family four: heuristics and metaheuristics
At the far end sit heuristics — rule-based and search-based methods, from simple greedy procedures to evolutionary and annealing-style searches. Their appeal is that they will produce something for almost any problem, however ugly, and they scale. Their weakness is that they offer no guarantee about how good that something is, and their quality can be quietly sensitive to tuning. Used with eyes open and validated honestly, heuristics earn their place, especially as a fast pass or a fallback. Used as a black box whose answers are never checked against ground truth, they are how firms end up trading on portfolios that nobody can defend.
The frontier: quantum approaches
Quantum and quantum-inspired methods get a great deal of attention for combinatorial problems, and the attention is not entirely unearned — certain selection and position-count problems map naturally onto formulations these machines are designed to explore. But the honest state of the field in 2026 is that quantum approaches are a research track, not a production engine for portfolio construction. The hardware is improving, the theory is genuine, and the day may come; today, anyone claiming a quantum tax-alpha miracle is selling a story, not a result. The credible posture is to keep a quantum lane open as research while being completely transparent that the production results come from classical computation.
Methods are means, not ends. What a buyer should care about is the outcome on their own data — speed, scale, after-tax dollars, and whether the answer holds up against an exact reference. That is exactly what a matched-workload pilot measures.
Request a pilot →So which one should you use?
The unsatisfying, correct answer is that production systems rarely live in a single family. The real engineering question is not "convex or integer or heuristic?" but "how do you route a given problem to the approach that fits it, hold a measurable quality bar, and do it fast enough and at enough scale to clear the whole book?" That routing question — matching problem shape to method, under a hard deadline, across a large and heterogeneous book — is where the genuinely hard and genuinely valuable engineering lives. The method names are table stakes; the system that orchestrates them under real-world constraints is the moat.
References & further reading
- H. Markowitz, "Portfolio Selection," The Journal of Finance, 1952 — the origin of the mean-variance framework.
- S. Boyd and L. Vandenberghe, Convex Optimization, Cambridge University Press, 2004 — the standard reference for convex methods.
- D. Bertsimas and J. Tsitsiklis, Introduction to Linear Optimization — for the integer-programming and combinatorial background.
- Asymmetry Computing, Reading solver benchmarks like an adversary.