A continual learning process is applied to a class of risk estimate-based algorithms and associated risk thresholds used for deciding when to initiate a handoff between different types of network connections that are available to a mobile device having telephony functionality. The process is implemented as a virtuous loop providing ongoing tuning and adjustment to improve call handoff algorithms and risk thresholds so that handoffs can be performed with the goals of minimizing dropped calls and unacceptable degradation in call quality and avoiding premature handoffs. Device characteristics, environmental context, connection measurements, and outcomes of call handoff decisions are crowd-sourced from a population of mobile devices into a cloud-based handoff decision enabling service. The service evaluates potentially usable handoff decision algorithms and risk thresholds against archived crowd-sourced data to determine how they would have performed in real world situations and delivers improved algorithms and risk thresholds to the mobile devices.