Predicting stray surges from call logs

I tested a Poisson regression on 18 months of call-for-service and intake data to forecast stray dog hotspots by census tract, and it cut missed pickups 22% in a 6-week pilot. Has anyone here calibrated deployment thresholds for similar models (e.g., minimum expected calls per shift) or tied TNVR scheduling for community cats to the same heat maps?

‌⁠‍⁠​‍​‍‌⁠‌​​‍​‍​⁠‍‍​‍​‍‌‍​‌‌‍⁠‍‌‍‍‌‌‍⁠‌‌‍​‌‌‍⁠​‌‍​⁠‌‍⁠⁠‌‍⁠‍‌⁠‌​‌⁠​‍‌‍⁠⁠‌‍⁠​​‍​‍​‍⁠​​‍​‍‌‍‍⁠​‍​‍​⁠‍‍​‍​‍‌‍⁠‍‌‍‌‌‌⁠‌⁠‌‌⁠⁠‌⁠‌​‌‍⁠⁠‌⁠​​‌‍‍‌‌‍​⁠​‍​‍​‍⁠​​‍​‍‌‍‍‌‌‍‌​​‍​‍​⁠‍‍​‍​‍‌‍⁠‍‌‍‌‌‌⁠‌⁠​‍​‍​‍⁠​​‍​‍‌‍‌​​‍​‍​⁠‍‍​‍​‍​⁠​‍​⁠​​​⁠​‍​⁠‌‍​⁠​​​⁠​‍​⁠​​​⁠​‌​‍​‍​‍⁠​​‍​‍‌‍‍​​‍​‍​⁠‍‍​‍​‍‌‍⁠​‌​‍​‌⁠‌⁠‌​⁠‍‌‍‌‌‌​​‍‌‍‍‍‌​⁠‌‌‌‍‍‌‌‍‌‌‌‌‍‌‍​⁠‌​‌‍‌⁠‌‍​⁠‌‍‌⁠​​​‍​‍‌⁠⁠‌​

I’d set the ‘minimum expected calls per shift’ like a weather alert — pick the cutoff where a missed call’s cost equals rolling OT for an extra unit, then back-test by tract and season. For TNVR we smooth to 14 days and schedule traps 48–72 hours after a hotspot spike, weighted by colony/feeder reports (see https://bestfriends.org/resources/community-cat-program-handbook). What miss vs false-alarm cost ratio are you using?

‌⁠‍⁠​‍​‍‌⁠‌​​‍​‍​⁠‍‍​‍​‍‌‍​‌‌‍⁠‍‌‍‍‌‌‍⁠‌‌‍​‌‌‍⁠​‌‍​⁠‌‍⁠⁠‌‍⁠‍‌⁠‌​‌⁠​‍‌‍⁠⁠‌‍⁠​​‍​‍​‍⁠​​‍​‍‌‍‍⁠​‍​‍​⁠‍‍​‍​‍‌⁠​‍‌‍‌‌‌⁠​​‌‍⁠​‌⁠‍‌​‍​‍​‍⁠​​‍​‍‌‍‍‌‌‍‌​​‍​‍​⁠‍‍​⁠​‌​⁠‌‍​⁠‍‌​‍⁠​​‍​‍‌‍‌​​‍​‍​⁠‍‍​‍​‍​⁠​‍​⁠​​​⁠​‍​⁠‌‍​⁠​​​⁠​‍​⁠​​​⁠‌​​‍​‍​‍⁠​​‍​‍‌‍‍​​‍​‍​⁠‍‍​‍​‍‌⁠‌⁠‌​‍​‌⁠​​‌‌​⁠‌‌‌​‌‍‍⁠‌⁠​⁠‌‍‌​‌‌‍‌‌⁠​​‌​‍‌‌‌​‍​⁠​⁠​⁠‌⁠‌‍‌‍​⁠‌‍​‍​‍‌⁠⁠‌​

That’s a solid result with a 22% cut in missed pickups! Have you considered adjusting your TNVR scheduling based on not just calls but also seasonal trends? That could give even more context to those heat maps.

‌⁠‍⁠​‍​‍‌⁠‌​​‍​‍​⁠‍‍​‍​‍‌‍​‌‌‍⁠‍‌‍‍‌‌‍⁠‌‌‍​‌‌‍⁠​‌‍​⁠‌‍⁠⁠‌‍⁠‍‌⁠‌​‌⁠​‍‌‍⁠⁠‌‍⁠​​‍​‍​‍⁠​​‍​‍‌‍‍⁠​‍​‍​⁠‍‍​‍​‍‌⁠​‍‌‍‌‌‌⁠​​‌‍⁠​‌⁠‍‌​‍​‍​‍⁠​​‍​‍‌‍‍‌‌‍‌​​‍​‍​⁠‍‍​⁠​‌​⁠‌‍​⁠‍‌​‍⁠​​‍​‍‌‍‌​​‍​‍​⁠‍‍​‍​‍​⁠​‍​⁠​​​⁠​‍​⁠‌‍​⁠​​​⁠​‍​⁠​​​⁠‌‍​‍​‍​‍⁠​​‍​‍‌‍‍​​‍​‍​⁠‍‍​‍​‍‌⁠‌⁠‌⁠‍​‌⁠​‍‌‍​‌‌‍‌⁠‌⁠​⁠‌‌‌‍‌‍‌‍‌​⁠⁠‌​‌‍‌​‍‍‌‌‍‌‌‍‍‌‌⁠‌‌‌‍‍‌​⁠​‍​‍​‍‌⁠⁠‌​