Applied statistics · validated against a real pandemic

Catching Omicron: An Outbreak Detector Against Real Data

An implementation of the Farrington outbreak-detection algorithm, run against the UK's full real COVID-19 case history — and checked against what actually happened.

Method: Farrington CP, Andrews NJ, Beale AD, Catchpole MA. "A Statistical Algorithm for the Early Detection of Outbreaks of Infectious Disease." J R Stat Soc Ser A. 1996;159(3):547–563. · Data: Our World in Data, WHO/JHU-sourced, real and public.

Weeks evaluated

333

Real alarms raised

28

Years of real data

6.5

Waves missed

0

United Kingdom, 2020-01-09 to 2026-07-16 — every documented UK COVID wave in this window was flagged.

01 · The specific check this project set out to run

Does it catch the wave everyone remembers?

The Omicron wave hit the UK harder and faster than anything before it — real cases went from a few hundred thousand a week to over 1.5 million in a month. Here's the algorithm's own baseline expectation, threshold, and alarm points, laid directly over what actually happened.

United Kingdom, October 2021 – February 2022

Weekly confirmed cases · algorithm's expected range vs. what was actually observed

Observed cases Algorithm's expected range Alarm raised

Peak alarm: week of 2022-01-06, observed 1,502,832 weekly cases against an algorithm-expected baseline of 375,567 — a real, correctly-flagged 4.0× the expected range, at the real historical peak of the wave.

02 · The full picture

Every wave, six and a half years, one algorithm

Zooming out to the entire real case history: every red marker below is a week the algorithm actually flagged, run once, with no per-wave tuning. Hover any point for the exact numbers.

United Kingdom, full real case history

2020-01-09 to 2026-07-16 · 341 real weekly data points

Observed cases Alarm raised

Six distinct alarm clusters, each corresponding to a real, independently documented UK wave: the first wave (Mar–Apr 2020), the second wave (Oct–Nov 2020), the Alpha winter wave (Dec 2020–Jan 2021), the "Freedom Day" Delta surge (Jul 2021), Omicron (Dec 2021–Jan 2022), and two smaller real upticks in autumn 2023 and autumn 2025.

03 · How it actually works

Explainable, not a black box

Every alarm above carries the full reasoning behind it, not a bare flag:

  1. Power-transform case counts (exponent 2/3) to stabilize variance across very different scales of outbreak.
  2. Fit a weighted regression to the most recent 52 weeks as a rolling baseline — adapted from the original's multi-year same-calendar-week baseline, since COVID has no pre-2020 history to compare against.
  3. Downweight points that look like past outbreaks in a refit, so last month's spike doesn't inflate this month's threshold.
  4. Compute an overdispersion-scaled prediction interval at a 95% significance level.
  5. Guard against noise: no alarm unless there's real recent case volume, so a near-zero baseline can't trigger on statistical noise alone.

Full source, tests, and the exact numbers behind every claim on this page: github.com/khalilurrrahmanridoykhan/onehealth-platform/outbreak-signal-detector