––° ––mph ingest HOLD

Haven Reach · synthetic · day 7

Home Research Rapid Epidemic Growth

Research · modeled study · synthetic metapopulation

Catching epidemic
rapid growth

A seed-locked ensemble on a fictional 48-county commuting graph. Rapid growth onset defined as a ≥2× rise in 7-day incidence with a nowcasted reproduction number remaining ≥1.3 — a threshold that separates local flicker from regional acceleration.

View on GitHub

48

Counties

Haven Reach metro

≥2×

RGO definition

per 7 days

1 d

Cadence

incidence nowcast

200

Seasons

seed-locked ensemble

The question

When does a respiratory pathogen on a commuting graph switch from local flicker to regional acceleration — and can we see it seven days early?

Hospital occupancy, school-closure timing, and staffing models all depend on a growth signal that is still local. When incidence doubles in a week and stays supercritical, every bed model, every absentee forecast, and every interruption rider priced on last week’s curve goes stale at once. The goal here is methodological: define a rapid-growth onset (RGO) that is the epidemiological analog of a ≥30 kt / 24 h wind jump, then test whether a nowcast on a commuting graph can flag it before occupancy moves.

Haven Reach is fictional. Populations, origin–destination flows, and linelists are constructed. Nothing on this page is a claim about a real outbreak, a real metro, or a published result. It is a modeled study — the same pipeline shape as the hurricane note, pointed at a different field.

Seven days

Ensemble draw 047 · days 0–7
West mill Airport Haven University Eastgate The Reach
Day 0 · university seed Local flicker · Rê ≈ 1.1
West mill Airport Haven University Eastgate The Reach
Day 7 · regional field RGO · incidence ≥2×

This is the jump the pipeline is built to flag. A university-county seed is still a local cluster on day 0. Seven days later the downtown, airport, and Eastgate hubs are lit — a ≥2× seven-day incidence gain with Rê remaining above 1.3, visible on the graph as a corridor, not a single patch.

7-day incidence ratio · nowcast vs occupancy

TanStack Charts · dashed rule is the ≥2× RGO threshold. Ensemble draw, not observations.

The graph

Schematic · gravity OD
West mill Airport Haven University Eastgate The Reach
Hub commuting edges
County node
RGO onset county

Nodes are synthetic counties; edges are gravity-model commuting flows. Rapid-growth onsets in the ensemble concentrate on hub counties — downtown Haven, the airport, the university, Eastgate — where between-patch mixing is high enough to turn a local cluster into a regional wave.

Method

01

Synthetic population & commuting OD

Forty-eight counties given constructed populations and a gravity origin–destination matrix. Hub weights (downtown, airport, university, port, Eastgate) set mixing; weekend and holiday schedules thin the same matrix. No census tract, no real metro.

02

Stochastic metapopulation SEIR

Each county is an SEIR compartment with between-patch force of infection from commuting. Transmission, latent, and infectious periods drawn once per season. Tau-leaping updates at daily cadence; hospital occupancy lagged from incidence by a fitted delay.

03

RGO flagging & nowcast

For each day, compute the 7-day incidence ratio and a nowcasted Rê. Observations where the ratio ≥ 2.0 and Rê remains ≥ 1.3 tagged as rapid-growth onset. Consecutive flags collapsed to a single event. Weekend reporting bias corrected before the ratio is taken.

04

Reproducibility

Two hundred seasons, seed-locked. Pipeline is a single notebook: constructed OD → ensemble → figures. Environment pinned via conda-lock. Figures generated with matplotlib; any researcher can rerun end-to-end from the in-repo synthetic files.

What the ensemble shows

~61%

of first RGO flags on hub counties

Downtown Haven, the airport, the university, and Eastgate account for most first onsets. Peripheral counties rarely lead; they follow the commuting corridors.

6–9 d

median nowcast lead vs occupancy

The RGO flag precedes the hospital occupancy upturn by about a week in the ensemble. That is the operational window the threshold is meant to buy.

~11 d

peak delay if top 8% of edges close

A counterfactual that zeros the heaviest commuting edges delays the regional incidence peak. It is a sensitivity result, not a policy recommendation.

Weekend bias

inflates false RGO if unadjusted

Raw seven-day ratios fire on reporting artifacts. The pipeline debiases weekday completeness before the threshold is applied; without that step, precision collapses.

RGO events per season · 20-draw slice

Season count
Linear trend
Recent seasons

Illustrative slice of a 200-season synthetic ensemble. Not observations.

RGO volume

Isometric · window × hub

Mean first RGO flags · three-dimensional

Left / right faces
Earlier windows
S16–20 top faces

Each column is a five-season window × hub cell. Height is mean first RGO flags in the ensemble. Downtown Haven leads every window; later seasons lift as amber caps. TanStack Charts, isometric projection of constructed ensemble counts.

Why it matters

Hospital capacity

Occupancy models trained on last week’s incidence miss the week the curve bends. An RGO flag is a staffing trigger: the lead time is the difference between a planned surge and a scramble.

School and workplace timing

Closure and remote-work decisions priced on county case counts arrive after commuting has already mixed the patches. A graph-aware threshold sees the corridor, not just the seed county.

Business interruption

Interruption riders and absentee forecasts assume a smooth epidemic curve. Rapid growth is a regime change. Pricing that ignores onset understates the week that actually costs.

Forecast model improvement

Identifying the mixing edges that reliably precede RGO gives nowcasting a target: resolve those corridors, and growth forecasts improve where occupancy is about to move.

Stack

Python 3.11 Runtime
numpy / pandas State & linelist
networkx Commuting graph
scipy Nowcast / Rê
matplotlib Figures
TanStack Charts Site figures

Data sources

Synthetic OD Gravity matrix
Constructed pops County weights
Ensemble linelist Primary
Occupancy lag Derived

Reproducibility

Pipeline is a single notebook from constructed OD to ensemble figures. No external outbreak feed.

Environment locked with conda-lock. All random seeds fixed. Re-runnable from the in-repo synthetic files.

johnathan-radojevich/rapid-epidemic-growth →

Next project

Decoding hurricane rapid intensification →
View on GitHub