Answers are grounded in this paper and cite their sources.
A phone that shows a 5G icon in a moving car does not keep it for long, and how long it keeps it has rarely been measured as a duration.This paper treats every stretch of New Radio service in public device-side traces from a commercial network as a survival-time observation, with trace ends handled as censoring.Kaplan-Meier curves, five parametric families, a Cox model and a discrete-time hazard regression describe the dwell time; logistic and gradient-boosted classifiers, validated with whole traces held out, then try to announce a fallback ten seconds ahead.On the road the median dwell is 34 s, against more than an hour and a half when the phone is still.A two-component exponential mixture fits best: a fresh connection is about eighteen times more likely to drop in the next second than one that has lasted a minute.Signal level, speed and recent instability move the hazard; the neighbour-cell gap does not.At an alarm duty of 10% the boosted model warns of 83% of fallbacks, and a map of earlier fallbacks improves it further, since they recur at the same places.Mean download rate halves across a fallback while the median barely moves.
No takes yet. Share an insight, caveat, or question.
Lobna Osman (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: