Public data on what's going around, for people, pets and farm animals. No accounts, no cookies. How it works

Methods

How the forecast works. And where it doesn't.

Public numbers in, a reading of this week and the next four out, each with a range and a reason to trust it or not. Engine: Vitals Forecast engine v0 (pooled damped trend + graph spread model)

01

Collect

Every public source is pulled on its own schedule and every pull is kept as dated as-of snapshots. A replay of any past week sees only what was public that week.

02

Nowcast

Data arrives late and gets revised. The nowcast estimates this week's value from what has come in so far, with a range. Where a county has no monitoring, we borrow from the nearest monitored areas and say so.

03

Forecast

One to four weeks ahead, each week as a likely value with 50% and 90% ranges, set so that over past seasons about half and about nine in ten outcomes fell inside them. Event threats, like a new confirmed H5N1 dairy herd in a state, get a probability instead.

Levels and trends. Measured against each place's own past.

A level of High means high for that place, compared with its own last two years of weekly values. Trend compares this week with last week: up more than 10% is rising, down more than 10% is falling, anything between is steady.

The rule the engine used for this build:

kind
cuts
levels
  • Very low
  • Low
  • Moderate
  • High
  • Very high
trend
rising_pct
10
falling_pct
-10
percentile_cuts
  • 0.2
  • 0.5
  • 0.75
  • 0.9
event_cuts
  • 0.05
  • 0.15
  • 0.35
  • 0.6
note
Per region: value cuts at the 20th, 50th, 75th and 90th percentiles of its own last 104 weeks.
cuts
858 regions

Two locked hindcasts. Run once on past seasons, with only the data public at the time, and published as they came out.

h5n1_dairy_2024REAL

H5N1 in US dairy herds, 2024: the model ranked states better than the baselines, but its probabilities ran too high, so on the locked primary score it did not beat them.

question
Which states get a new confirmed dairy herd in the next 4 weeks? 40 weekly forecasts, Apr-Dec 2024, 48 states, as-of data only.
primary_brier
vitals_v0
0.06346
persistence
0.05976
neighbour_spread
0.06006
primary_auc
vitals_v0
0.8729
persistence
0.7821
neighbour_spread
0.7819
verdict_vs_persistence
diff
0.0037
ci95
  • -0.00223
  • 0.00985
verdict
not shown better
verdict_vs_neighbour_spread
diff
0.0034
ci95
  • -0.00217
  • 0.00944
verdict
not shown better
states_not_yet_affected_auc
vitals_v0
0.6886
persistence
0.4838
neighbour_spread
0.4939
misses
  • NV 2024-11-11: forecast 1%, herd confirmed
  • UT 2024-10-07: forecast 1%, herd confirmed
  • NV 2024-11-18: forecast 2%, herd confirmed
  • UT 2024-10-21: forecast 2%, herd confirmed
  • UT 2024-10-14: forecast 2%, herd confirmed
  • NV 2024-12-02: forecast 2%, herd confirmed
notes
10 of 40 origins used the plan's lag-proxy (no APHIS snapshot within 14 days). Cattle movement is a gravity-model ESTIMATE.
files
trials/vitals_forecast/results-h5n1-dairy-2024-v1.json
flu_2023_24REAL

US flu hospital admissions, 2023-24: Vitals scored 1.28x the CDC FluSight ensemble's weighted interval score (lower is better; worse).

question
Weekly confirmed flu admissions by state, 0-3 weeks ahead, every 2023-24 FluSight reference date, as-of data only.
rel_wis_vs_ensemble
rel_wis
1.2831
ci95
  • 1.0816
  • 1.5093
verdict
worse
rel_wis_vs_baseline
rel_wis
0.9487
ci95
  • 0.7508
  • 1.2142
verdict
not distinguishable
ensemble_rel_wis_vs_baseline
0.7394
coverage_95
vitals
50
0.532
95
0.969
ensemble
50
0.49
95
0.92
baseline
50
0.285
95
0.896
misses
  • HI: 1.66x the ensemble
  • NV: 1.60x the ensemble
  • CO: 1.60x the ensemble
  • MN: 1.60x the ensemble
  • ME: 1.57x the ensemble
files
trials/vitals_forecast/results-flu-2023-24-v1.json

Limits. Read these before you act on a number.

  • Wastewater shows what's in a community's sewage, not how many people are ill. Sites change, and small sites are noisy.
  • Hospital data covers people sick enough to be admitted, and arrives a week or more late.
  • Animal reports undercount. Few cats are tested, and herd and flock confirmations depend on who looks and when.
  • No free public source tracks kennel cough or parvo, so we show "No public data" rather than guess.
  • Forecasts are least sure at turning points: the start of a season, a peak, a new strain.
  • What-if runs on the operator map are SIMULATED scenarios on a simple spread model. They aren't forecasts.
  • None of this is about any one person or animal. It is not medical or veterinary advice.

Sources. Every one, with its licence and the credit line we show.

SourceThreatsFinest levelUpdatesLicenceCommercial useCredit line we show
HPAI confirmed cases in livestock (dairy herds)
USDA APHIS (history via CDC mirrors and the Internet Archive) · aphis_dairy
h5n1statedailyUS-PDYesSource: USDA APHIS. Not endorsed by USDA.
APHIS HPAI detections in mammals
USDA APHIS · aphis_mammals
h5n1county, stateirregularUS-PDYesSource: USDA APHIS
APHIS confirmed HPAI in commercial and backyard flocks
USDA APHIS · aphis_poultry
h5n1county, statedailyUS-PDYesSource: USDA APHIS
APHIS HPAI detections in wild birds
USDA APHIS · aphis_wildbird
h5n1county, stateirregularUS-PDYesSource: USDA APHIS
Interstate cattle shipment shares (gravity-model ESTIMATE)
Vitals Forecast (derived from USDA NASS and US Census) · cattle_movement
h5n1stateannualUS-PDYesEstimated from USDA NASS inventories and US Census population centres
CDC West Nile virus (ArboNET and NNDSS weekly)
CDC Division of Vector-Borne Diseases / NNDSS · cdc_arbonet
wnvcounty, stateweeklyUS-PDYesSource: CDC ArboNET; CDC NNDSS
NHSN Weekly Hospital Respiratory Data (HRD) by Jurisdiction
U.S. Centers for Disease Control and Prevention (NHSN) · cdc_nhsn
covid, flu, rsvstate, countryweeklyUS-PDYesSource: U.S. Centers for Disease Control and Prevention (CDC). Not endorsed by CDC.
CDC National Wastewater Surveillance System (NWSS)
U.S. Centers for Disease Control and Prevention · cdc_nwss
covid, flu, rsv, h5n1, measles, mpoxcounty, state, countryweeklyUS-PDYesSource: U.S. Centers for Disease Control and Prevention (CDC). Not endorsed by CDC.
CDC Lyme disease cases by county
CDC Division of Vector-Borne Diseases · cdc_tick
tickcounty, stateannualUS-PDYesSource: CDC, Lyme disease surveillance data
ECDC ERVISS: European Respiratory Virus Surveillance Summary
European Centre for Disease Prevention and Control · ecdc_erviss
resp, flu, rsv, covidcountryweeklyCC-BY-4.0YesDataset provided by ECDC based on data provided by public health authorities (ERVISS, erviss.org); adapted by Vitals Forecast.
EPA CyAN satellite cyanobacteria index (largest lakes per state)
US EPA (with NASA, NOAA, USGS) · epa_cyan
habscounty, stateweeklyUS-PDYesSource: US EPA Cyanobacteria Assessment Network (CyAN)
NOAA nClimGrid-Daily state temperatures and NWS 7-day outlook
NOAA NCEI / National Weather Service · noaa_weather
heatstatedailyUS-PDYesSource: NOAA NCEI nClimGrid-Daily; NOAA National Weather Service
USDA NASS Cattle (January 1 inventory by state)
USDA National Agricultural Statistics Service · usda_nass
h5n1country, stateannualUS-PDYesSource: USDA NASS, Cattle report

Tags. On every number.

REAL

Published by the source, as published (after unit conversion).

ESTIMATE

Worked out by Vitals from public data: nowcasts, forecasts, borrowed values for unmonitored counties.

SIMULATED

Made up: test data in a DEV build, or a what-if scenario. Never mixed into a real forecast.