Kinesthetic

Lead story · Drought in the Horn of Africa

It warned before all seven major droughts. A median seven months ahead.

We took every major Horn of Africa drought since 1996, fixed the date each crisis was declared from ReliefWeb, OCHA and FEWS NET records before any scoring began, and then asked one question: looking only at public data as it stood at the time, would Prism have warned?

7 of 7major droughts warned of before the crisis was declared.
7 monthsmedian warning ahead of the declaration. The range was 4 to 11 months.
0.40skill score, far beyond chance (p = 0.002).
86rainy seasons backtested, 1982 to 2024.

How it was tested

The seven drought dates were fixed from the public record first, so nothing could be tuned to fit them afterwards.

It works from public data only. Nothing private, and nothing of anyone's.

A warning counts only if it came before the crisis was declared. Seven droughts, seven warnings, a median of seven months in hand.

What this does not show

  • About half of all warnings were false alarms. A warning means "watch this", not "a drought is coming".
  • FEWS NET was earlier in 3 recent cases. The US famine early warning network, with far more inputs, beat us to it three times.
  • Seven droughts is a small sample. A skill score built on seven events is a strong signal, not a settled fact.
  • It is a starting point. Something to sit beside the existing warning systems, not to replace them.

The same idea, elsewhere

Read the record, say so early. It travels.

Cancer, from the notes a practice already holds

44%
of cancers flagged a median 3.2 months before diagnosis, following the NHS's own NG12 referral guidance. Bowel cancer: 84% flagged, 5.7 months early.

In testing on simulated patient records; not yet validated on real records. These are not clinical results.

The result in Medical Prism →

Leaks, from a water network's own readings

16 of 23
hidden leaks found in a town's network. That matched the best published competition entries.

A benchmark result on one town's network, scored against the published entries for the same task.

Household water-treatment plans

99 of 99
numbers checked against WHO guidelines came back right.

It never calls water safe. It says which tests are still missing before anyone could.

What it could watch next

Aims, not results.

None of these is done. Each is a place where the record is public, the pattern repeats, and months of warning would change what people can do.

FamineWe want to carry the drought work forward to food security, beside the systems that already do this.Aim
OutbreaksWe want to read public health surveillance for the signs that came before past outbreaks.Aim
HeatWe want to give towns weeks of notice before the heat that fills hospitals.Aim
Medicine shortagesWe want to spot a shortage in the supply record before the pharmacy counter does.Aim

Work with us

Have a record you think holds a warning?