Why semi-optimal Sharpe ratio allocation has become the working practitioner’s compromise
Modern portfolio theory is one of those disciplines whose intellectual architecture was largely settled by 1959 and whose practical execution has been quietly revised ever since. The mean-variance framework Harry Markowitz produced sits in nearly every introductory finance textbook in roughly the form he gave it, and the Sharpe ratio that grew out of the same conceptual ground has become the default scorecard for asset managers, hedge funds, and trading desks alike. Read the practitioner literature more closely, however, and the field has spent six decades addressing the points at which the original framework either assumes too much, requires inputs that cannot be reliably estimated, or produces optimal portfolios that no working allocator would actually hold. The honest version of the discipline now lives somewhere between the theoretical optimum and the practical compromises required to make allocation work in conditions the original theory did not anticipate.
Giulio Occhionero, Head of Quantitative Research and Development at IRH Global Trading and a published researcher on quantitative finance topics, has worked extensively in this territory. His SSRN paper on what he terms semi-optimal Sharpe ratio allocation is part of a broader effort to define an analytically tractable middle ground between the unconstrained mean-variance optimum, which depends on inputs that are notoriously unstable, and the equal-weight or heuristic allocations that practitioners often fall back to when the optimization machinery becomes unreliable.
What has portfolio theory actually learned in the time between Markowitz’s original framework and current practice?
Several things, none of which the standard textbook treatment fully reflects.
The most consequential change is the recognition that the inputs to mean-variance optimization, expected returns and covariances, are estimated with errors large enough that the optimizer routinely produces allocations more extreme than the underlying analysis supports. The polite name for this in the academic literature is estimation error. The blunter name, common in practitioner conversation, is that the optimizer is a noise amplifier. Small changes in input assumptions produce large changes in output weights, and the output weights themselves are frequently dominated by long-short positions of a size that no risk committee would approve. The discipline’s response has been a long migration toward constrained optimization, robust estimation, shrinkage estimators, and outright resampling, all of which represent admissions that the original framework, applied naively, does not survive contact with real data.
A second change concerns what the optimization is actually trying to do. Markowitz’s framework optimizes a single-period mean-variance objective. Real allocators are managing multi-period processes with path-dependent constraints, leverage limits, drawdown tolerances, and a horizon that is rarely a clean single period. The Sharpe ratio, used as a scorecard, hides much of this complexity, but the working practice of allocation has had to confront it. Occhionero’s semi-optimal framework is one response to that confrontation. By minimizing portfolio variance subject to the structure that allocation weights actually need to satisfy, the framework produces allocations that approximate the Sharpe-optimal solution while remaining stable enough to deploy in conditions the unconstrained optimizer cannot handle.
A third change, less acknowledged in standard treatments, concerns the role of correlation in portfolio construction. The original framework treats correlation as an input to be estimated and used. The more careful version of the discipline treats correlation as a screening criterion as well. Highly correlated strategies, above some threshold, are mathematically prone to producing negative allocation coefficients in optimized portfolios, which translate into long-short positions that may be neither intended nor advisable. The practical response, which Occhionero’s framework formalizes, is to drop strategies whose correlations exceed the threshold before optimization rather than after, allowing the diversification benefit to come from genuinely independent return sources rather than from optimizer artifacts.
A fourth change concerns what the framework can do that the original could not. Applied to baskets of correlated assets, the semi-optimal approach produces synthetic instruments, effectively constructed ETFs, whose statistical properties are tighter than any of the underlying components. Applied to US equity futures, the S&P 500, Nasdaq, and Dow Jones, the construction yields a stationary basket whose variance is dramatically lower than any of the individual contracts. Applied to US Treasury futures, where component correlations exceed 98 percent, the construction yields a fixed-income synthetic with properties no individual contract delivers. The framework is, in this sense, a constructive tool as well as an analytical one. It produces objects that did not previously exist as tradable units and that behave better than the components they are built from.
These developments are not abstract. The semi-optimal framework, applied to a portfolio of trading strategies rather than a portfolio of assets, allows the allocator to treat strategies as investable units in their own right, with their own Sharpe ratios, their own correlations to one another, and their own contributions to portfolio-level performance. The result is a strategy allocation discipline that is structurally similar to asset allocation but better suited to the actual question a quantitative fund manager faces, which is rarely how much of one stock to hold and almost always how much of one strategy to deploy.
The broader point, for institutional allocators and quantitative researchers whose work depends on portfolio optimization producing usable answers, is that the discipline’s working practice has moved a considerable distance from the textbook treatment of mean-variance optimization, and the movement has been driven by the recognition that optimal in theory and deployable in practice are not always the same thing. Semi-optimal frameworks, of which Occhionero’s is one, represent the honest compromise the discipline has been quietly working toward for decades. They do not claim to solve the optimization problem in its strongest form. They do claim to solve a usable version of it, with inputs that can be estimated, outputs that can be held, and properties that survive the kind of stress conditions under which the unconstrained optimum tends to fail.
That is the standard the working practice has been converging on. It is also, on any honest reading, the standard that distinguishes allocation work that holds up from allocation work that does not.
Giulio Occhionero is Head of Quantitative Research and Development at IRH Global Trading. His research on semi-optimal Sharpe ratio allocation and related topics in quantitative finance is published on SSRN.
