Methodology
How Nyman Intelligence Biosecurity works
Questions addressed
Official case and death counts are authoritative and they arrive with a delay. Between situation reports a response team is working from a figure that can be several days old. This platform reads public reporting continuously and answers three questions in that gap.
- •Where does this outbreak stand today? A reconstructed curve and an estimate of the current count, with a range around it.
- •What should I be watching that has no name yet? A regional watch list that includes suspected clusters and clusters with no confirmed pathogen, ranked by how much attention each one warrants.
- •What is happening around the case count? Events showing the response under pressure, and events at national borders, each one traceable to the reporting behind it.
Data collection and processing
Four stages run continuously. Each one produces something the next stage uses, and every event on the platform can be followed back through them to the reporting it came from. Select a stage to read what it does.
Official bulletins, expert and scientific reporting, humanitarian sources, and regional and local news in the languages of the countries concerned. More than 250 distinct reporting streams have contributed so far, with French and Swahili sources alongside English. Collection runs through the day, so the record keeps growing between situation reports.
Estimation and accuracy
The curve is rebuilt from figures published in open reporting. It takes no official figures as input. The estimate of today’s count anchors on the most recent official figure and projects it forward at the rate the open record shows, and a short forecast extends that. Both come with a range.
The accuracy is measured, not asserted. In the established phase, with the last official figure a week old, the median error on the case count is 3.9 per cent, against 12.6 per cent for extending the official trend: about three times closer.
The advantage is concentrated where the official trend turns. Measured across every horizon from one to ten days, the median case error at those turning points is 13.6 per cent for the estimate and 27.0 per cent for the official trend. Where an outbreak is steady the official trend is already accurate and the estimate stays close to it. Every figure is scored out of sample, so at each past date the model saw only what existed on that date.
The full method is public: Nowcasting outbreak case and death counts from open-source reporting, medRxiv.
Source authority
Every outlet is scored continuously on how often its reporting matches what is eventually confirmed, how often it reports first, and how many events it has contributed to. That scoring is what the ranking rests on, so an event several independent outlets confirm counts for more than one outlet reporting alone. Each source’s record is on its own page, so the weighting is open to inspection.
The watch list, the curve and the estimate are open to everyone. Pro adds the event record behind every figure, the source profiles, and alerts you can set.
Institutional deployments, additional sources and custom models are arranged separately. See Access.