Storm Events
639,465 events. A lesson in instrument artifacts.
Decision Chain: Storm Events
Ingest NOAA Storm Events database. 639,465 events with hour-level timestamps.
639,465 events ingested. Hour-level precision. 48 storm types.
Fit self-exciting point process to all storm event timestamps.
Branching ratio 0.670. 211,024 independent (33.0%), 428,441 triggered (67.0%).
Check for structural breaks from changes in how storms are recorded.
1996 type expansion: categories grew from 3 to 48 storm types. This is an instrument change, not a weather change.
Analyze what fraction of the triggered events come from post-1996 storm types that didn't exist in the original dataset.
56% of Hawkes-triggered events come from post-1996 types. True physical clustering using only original 3 types would be meaningfully lower.
Branching ratio inflated by instrument expansion
The 0.670 branching ratio includes a massive instrument artifact: the 1996 expansion from 3 to 48 storm types created artificial event clustering. Over half (56%) of the Hawkes-triggered events are from categories that didn't exist before 1996. An instrument-corrected value using only the original storm types would show lower self-excitation. Some genuine physical clustering exists (storms do trigger other storms), but the measured value overstates it.