ADOWIP Optimizes Time-Series Adaptation with Budget.

Online time-series forecasting systems often face a dilemma: labels for predictions are only available after a delay, and every adaptation step consumes limited computational resources. Traditional approaches often adapt at every opportunity, which may not be optimal. This research focuses on *when* an online learner should update, rather than just *how* to adapt. The paper introduces ADOWIP (Adapt Only When It Pays), a residual-adapter framework designed for delayed online time-series adaptation. ADOWIP features sealed delay queues, precise budget accounting, and auditable update telemetry. Its core mechanism is an observed decision-loss priority gate, which triggers an update only after feedback is revealed, when the downstream loss (optionally penalized by prediction MSE) exceeds a calibrated empirical quantile, and sufficient budget remains. The authors prove hard-budget feasibility, projected-OGD regret for a convex linear accepted-update subproblem, and provide stability and conditional finite-sample gate-selection statements. Experiments on public ETT capacity-planning tasks show that a frozen calibration/evaluation split selects a gate that lowers held-out decision loss compared to always-on, fixed-period, and drift-triggered exact-update baselines, all under matched compute. Further tests on UCI Bike and Capital Bikeshare datasets also show improvements, though probe-based and finance experiments were negative, indicating scope limitations.