Judging By The Cover: Recoverable Signal of a Hidden Hardware Transition in Barcelona's Bicing E-Bike Fleet

This study examines how much of an internal hardware difference is externally recoverable during Barcelona's Bicing fleet sensor transition, in which cadence-sensor e-bikes are being progressively replaced with torque-sensor units. Around 300 e-bikes were assessed between January and April 2026 across 15 standardized field variables and classified into two performance types based on assistance behavior: Type A (cadence sensor, low-effort activation) and Type B (torque sensor, effort-requiring activation). Mann-Whitney U tests with rank-biserial effect sizes identified Inventory Tag — an ordered production-batch identifier visible on each bike — as the strongest signal, robust across collection subgroups, and Pedaling Rate (smoothness of backward pedaling) as the strongest behavioral signal, retaining independent effect after residualizing on production batch and collection conditions. A two-step rule combining these features reached 71.1% accuracy (±5.4%) against a 61.9% majority-class baseline (+9.2pp) under 10× repeated stratified 5-fold cross-validation, with Logistic Regression, Decision Tree, and XGBoost converging on a 68–72% band. An ablation removing all production-identifier features showed that behavioral signals in combination still recover up to 4-5 percentage points above baseline, though in isolation neither behavioral feature separates the types — confirming the signal is genuinely independent of the production identifier yet conditional on it. Findings suggest the externally recoverable signal decomposes into a dominant production-identifier component and a smaller behavioral residual, with within-batch variation appearing to limit further gains at this sample size.