By Vacslav Glukhov at ItoFlow: Starting with Harry Markowitz and continuing through Black-Litterman and numerous more recent works, the central academic problem of portfolio management has remained largely unchanged: optimizing allocations by maximizing the expected future return, penalized for risk.

In its simplest form, this amounts to choosing a portfolio allocation vector under a single empirical model, optimized one step forward. In empirical studies, this myopic portfolio is rolled through time with historical data, and its outcomes are compared against benchmarks.

Extensions such as Bayesian priors and shrinkage [1], resampling [2], empirically grounded heuristics [3], and more recent work emphasizing robust pipelines [4], and a crop of machine learning and reinforcement learning work too numerous to select one representative, primarily refine the model layer.

They polish covariance estimations, improve the extraction of latent structure (e.g., factors), or stabilize allocations. The object itself remains largely the same. There are, of course, important but less central strands. Long-horizon frameworks [5–7], robust control and model uncertainty [8] and [9], structural breaks [10], and transaction cost aware multiperiod models with parameter uncertainty [11] broaden the setting. These developments are welcome. Yet, they tend to retain the same underlying structure: optimizing expected outcomes within a specified model. And they often remain peripheral to practice.

It is almost self-evident that while portfolio management practice does not reject optimization outright, it simply does not treat it as central.

Why?

A plausible explanation is that portfolio management is embedded in an institutional setting that prioritizes survival.

Once survival becomes the primary concern, optimality is understandably downgraded. What takes its place is robustness. Investment catastrophes, sometimes total wipeouts, often follow from allocations finely tuned to a particular regime. Optimality begets fragility: within a purely empirical optimization and selection framework, fragility is a natural outcome unless explicitly constrained.

An investment firm must survive the possibility that its underlying assumptions fail altogether. Structural breaks, whether rare or frequent, are a defining feature of financial markets. Fragile allocations and organizations tend to fail almost inavoidably at these points.

Next, survival shifts attention from point-in-time allocations to trajectories. Rather than seeking a myopic optimum, the firm designs, or more often develops through trial and error, a survival-and-prosper policy, and operates with an appropriate strategy in place. Certain trajectories must be avoided, even at a high cost. Thus, constraints and mandates are embedded in the policy. In theory, constraints are often secondary. In practice, they are vital.

Time further complicates matters. The farther we look into the future, the less confident we become – the fog of uncertainty thickens with distance. Discounting is the standard, albeit crude, tool for dealing with this in theory, but in practice confidence decay may appear in other forms: reduced forward reliance on predictive signals, pull toward structurally robust allocations, and heuristic downscaling of returns and upscaling of risks across horizons.

Markets are complex, open, and evolving systems. Complexity and evolution limit the usefulness of purely empirical data-centric approaches.

The increasing reliance on machine learning introduces its own constraints: models that are expensive to train are slow to adapt and struggle in a setting where the underlying process is shifting.

In academic portfolio theory, models are central. Tractability, neat closed-form solutions, and clear stylized results are valued. In practice, models are treated with a degree of skepticism. Once transaction costs, regime uncertainty, model uncertainty, and learning dynamics are folded into the mix, conclusions become messy. And messy results are difficult to publish.

I am not suggesting that this divergence reflects pathologies in either approach. Theory and practice simply address different problems.

Academia asks: how can expected utility/return be improved within a given model and dataset?

Practice asks: how can an investment policy be constructed that remains viable under stress and uncertainty?

The difference is not subtle. A method that improves Sharpe at the cost of higher turnover or increased tail sensitivity may be rejected in practice: an approach that slightly reduces returns while stabilizing behavior and reducing fragility may be preferred.

Similarly, academia shrugs when practitioners rely on informal, heuristic, rule-of-thumb adjustments, such as volatility scaling rules that presumably encapsulate confidence decay. Academia may instead favor a clean, model-consistent result derived from assumptions that easily conform to reviewers’ expectations.

This is not to suggest that practice can do without academia or vice versa.

On the contrary, practice often benefits from formal ideas developed elsewhere. Robust sequential control, multiobjective decision-making, and operations under deep uncertainty – these methods are well studied in other engineering domains and remain underexploited in portfolio construction.

At the same time, there may be room for academic work to broaden its scope.

Incorporating epistemic humility into market modeling, explicitly accounting for changes in confidence, understanding the origins of informal practices, treating heuristics as legitimate approximations, and studying investment as an organizational and survival problem, rather than optimization, may open new directions.

Whether that leads to something both interesting and useful – let’s see.