Title: Interpreting Worker Utility as Hysteresis: A Preisach Framework for Transaction Acceptance in Gig Economies

Abstract:

Worker utility remains unobservable; we only witness its outcomes. Since each gig interaction results in a binary signal—either an acceptance or a rejection—this structural constraint suggests that the Preisach hysteresis model serves as the most appropriate representation for latent worker preferences. The Preisach operator characterizes aggregate output by integrating over a population of binary threshold elements, a mathematical structure that naturally arises when a heterogeneous workforce possesses distinct, private acceptance wages.

To estimate two distinct latent utility surfaces—acceptance utility, denoted as $U_1(X)$, and rejection utility, $U_0(X)$—we employ a dual-output neural network. This architecture features shared layers (256 units down to 128) and utilizes margin loss to enforce the condition that $U_1 \geq U_0$. The classification task is simplified by calculating the Preisach gap, defined as $U_1(X) - U_0(X)$, which is then fed into an XGBoost classifier alongside clip-stabilized price-to-threshold encodings.

Testing this pipeline on a dataset of 36,891 gig transactions yielded a Jaccard score of 0.827 and a ROC AUC of 0.799. Notably, the price-to-threshold encoding contributed an additional 11.0 percentage points to the AUC compared to using raw utility features alone. The model validates the directional asymmetry predicted by hysteresis theory: reductions in price suppress completion rates more significantly than equivalent price increases boost them.

When applied to the entire dataset, the model’s recommendations achieve a dual objective: lowering the total wage bill by 21.3% while simultaneously raising the expected fill rate by 9.7 percentage points. Specifically, for 74.2% of transactions where the probability of acceptance ($P(\text{accept})$) already surpasses 0.80, reducing the wage maintains the probability above the acceptance threshold (with a mean post-reduction probability of 0.972), thereby generating cost savings with a median of 31%. For the remaining 25.4% of transactions, a median wage increase of 7% recovers an additional 43 percentage points in acceptance rates. Crucially, a model lacking an explicit indifference zone would be unable to execute both strategies concurrently.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC