Forecast residuals for the diagnosis of hidden traffic events: the ranking of normal models depends on the alarm rule

Hidden traffic events such as lane closures, demand surges, signal faults and sensor failures are commonly detected by forecasting normal traffic and thresholding the residual. We ask what the two stages of such a detector should condition on: the normal model that produces the residual, and the rule that reads it over sensors and time. The study uses a simulated signalised corridor in which every event episode has an event-free twin sharing all random variables, so that the event's effect on each sensor is known exactly, and radar measurements from Interstate 24 with a located event log. The central finding is that forecasting accuracy alone does not determine detection quality: the ranking of normal models depends on the residual score and the alarm rule. Under an instantaneous threshold a forecaster that conditions on recent observations predicts the event it is meant to expose. On the corridor the seasonal prior is not improved upon by the selected hybrid under this score, while under a rule that integrates the residual the accurate penalised autoregression decisively becomes the best linear pairing, at a longer delay and a longer time in alarm; the freeway ordering does not reverse. The relation between accuracy and detection differs between model families under one rule, and a nonlinear network and a zero-shot foundation model are poor inputs under either rule for different reasons. Two results of narrower scope follow: adding the nearest sensor to the detector's input raises recall substantially in this benchmark, and the normal model that is best for detection under an instantaneous rule is the worst among the tested strategies for reconstructing the event-free counterfactual. The generator, configurations, splits, full sweep tables and scripts accompany the submission; the corpus and checkpoints are to be released on publication.