Skip to content

Rolling Hotspot Recalibration

Recalibration process — instantiates Vulnerability Hotspot Mapping and Hardening

Re-scores and re-ranks the hotspot map on a fixed cadence against what actually happened, so the map tracks a moving risk landscape instead of freezing on its first version.

A hotspot map decays: risk drivers shift, yesterday's fix worked, new units appear that were never scored. Rolling Hotspot Recalibration is the scheduled process that re-runs the mapping on a cadence and checks the previous map against what actually happened — which flagged hotspots produced incidents, which quiet units surprised everyone — then updates scores, weights, and rankings. Its defining move is to close the loop between predicted hotspots and realized incidents, so the model learns instead of ossifying. Where sentinel monitoring watches a few sites continuously, recalibration periodically rebuilds and revalidates the whole map.

Example

A city keeps a map of traffic crash "black spots" to target enforcement and road engineering. Built once, it slowly goes wrong: a junction gets a roundabout and stops being dangerous, while a newly widened road turns lethal. Rolling Recalibration re-scores the map each year and backcasts — laying last year's predicted black spots over last year's actual crashes. Two predicted hotspots had no crashes (over-flagged); three bad crash sites were not on the map at all (missed). Those errors feed the update: weights adjusted, the fixed junction retired from the list, the new road added. Crucially, the review distinguishes a site that quieted because the fix worked from one that quieted by chance — because a black spot flagged on one unusually bad year often calms on its own (regression to the mean), and crediting the fix for that reversion is how budgets get wasted.[1]

How it works

  • Recalibrate on a cadence. Re-score every unit on a fixed schedule with fresh data, so ranks reflect the current landscape rather than the one at launch.
  • Backcast against realized incidents. Overlay the prior map on what actually occurred and score hits, misses, and false alarms — the errors are the signal that corrects the weights.
  • Separate real change from noise. Distinguish drops caused by interventions from regression to the mean before retiring a hotspot or crediting a fix.

Tuning parameters

  • Cadence — monthly, quarterly, annual. Faster tracks change but chases noise and costs effort; slower is stable but stale.
  • Backcast window — how much incident history each recalibration weighs. Longer is stable, shorter is responsive to recent shifts.
  • Update aggressiveness — how far a single cycle's misses are allowed to move the weights. Aggressive adapts fast but overfits the last period.
  • Retirement rule — how much sustained quiet, and how much evidence that it is the fix rather than chance, before a hotspot leaves the list.

When it helps, and when it misleads

Its strength is that it keeps the map honest against reality and turns each cycle's misses into a better model, catching both hotspots that were fixed and ones that newly emerged.

Its signature trap is regression to the mean: a site flagged on a spike reverts on its own, and a naive before/after comparison credits the intervention for a drop that would have happened anyway.[1] Recalibrating too aggressively also chases noise, reshuffling ranks on random variation and destabilizing the very priorities it should steady. The discipline is to use a control or an expected-reversion baseline before crediting any fix, and to match the cadence to how fast the risk actually moves.

How it implements the components

Rolling Hotspot Recalibration fills the upkeep-and-validation components — the ones that keep the map current and check it against reality:

  • maintenance_and_recalibration_cadence — it defines and runs the scheduled re-scoring that keeps the map from freezing.
  • historical_incident_backcast — it validates the prior map against realized incidents, using the misses to correct weights and rankings.

It does not watch a few sites continuously between cycles (sentinel_monitoring_panel, Sentinel Site Monitoring) and it does not stress the current map against hypothetical futures (scenario_stress_overlay, Hotspot Tabletop Stress Test).

References

[1] Regression to the mean — extreme observations tend to be followed by ones closer to the average for statistical reasons alone. A hotspot identified on an unusually bad period will often improve on its own, so a before/after comparison without a control systematically over-credits whatever was done in between. ↩