Machine learning has substantially expanded optimisation capacity in financial systems, yet persistent phenomena—alpha compression, strategy crowding, and instability under regime shifts—remain unresolved. This article argues that these effects reflect a structural constraint not captured by existing accounts: optimisation under fixed evaluative criteria. Financial systems optimise over strategies while holding constant the criteria that define performance, inducing evaluative compression—a contraction in the set of strategies admissibly superior under shared definitions of risk and return. The article introduces a distinction between first-order failure (performance degradation under a fixed objective) and second-order failure (inadequacy of the objective itself), showing that contemporary systems are designed to address the former but not the latter. Analysis of the 2007 quant equity crisis demonstrates how shared evaluative criteria generate structural coupling across strategies, producing correlated losses that recalibration alone cannot resolve. To address this limitation, the article proposes evaluative operation, a minimal extension in which performance criteria become admissible objects of constrained, signal-triggered transformation. It develops evaluative accountability as the corresponding governance requirement and identifies risks such as evaluative arbitrage. The central implication is that, as machine learning scales in finance, the primary constraint is no longer optimisation efficiency but the structure of the criteria under which optimisation proceeds.
Peter Kahl (Sun,) studied this question.