Mercalli

Model card

Two models sit behind the forecasts: one predicts how long an open incident will last, the other which services are likely to follow an incident at a provider they depend on. This card says how they were trained and how accurate they are.

What is predicted

For each open incident, the time left until the vendor marks it resolved, as three percentiles: the 10th, the 50th (the median) and the 90th. The range from the 10th to the 90th is an 80% interval. Each forecast uses only what was known when it was made: the incident's phase, its updates so far, its wording, the service's own history and the time of day.

The intervals are widened by conformal calibration on a later period the model never trained on, so that they hold their stated coverage on new incidents.

How intensity is measured

A single reading of how disrupted things are right now, on a ten step scale named after the Mercalli scale for how strongly an earthquake is felt.

The score adds up the open incidents that are minor or worse and less than a week old, weighted by impact, with extra weight for incidents at providers that many other services depend on. Advisories, maintenance and incidents a vendor has left open for over a week do not count.

The level says how unusual the score is: each threshold is the score that a given share of past hours stayed below, computed from a year of history. Until that history is in place the thresholds are provisional, and the reading says so.

The thresholds will be listed here once the model card is published.

Limitations

  • The forecast is of when the vendor marks the incident resolved, which can lag the real recovery.
  • Vendors sometimes post late and backdate the start, or post only after an incident is over. Those incidents cannot be forecast while open.
  • Some incidents disappear from a page without a resolution time. Their end is when Mercalli noticed, and is marked as inferred.
  • Services with little history borrow from their category, so their forecasts are less specific.
  • The blast radius model sees only public signals: the census, past incident reports and past co-failures. A private dependency it cannot see will not show up.