For light-rail transit, shared routes (LRV + automobiles) and traffic signals mean that arrivals are not predictable. Where shared-roadway and dedicated-route segments exist (say, with San Francisco's Muni Metro), hold times must be built in for the dedicated-route travel segments. Train arrivals are highly unpredictable. A level below Muni, on Bart, rush-hour headways through the San Francisco subway are 4 minutes. That's close to the theoretical maximum Alon gives of two minutes.
For conventional heavy rail, boarding / debarking delays, equipment failures, medical emergencies, police activity, level-grade collisions, and other factors, may all lead to service delays.
Signalling matters mostly in that it both doesn't introduce new delays into the system, and that it can compensate for delays originating elsewhere.
Reliability overall, as discussed in the article, is largely based around signalling to manage flows.