Every method, on the record
Every method we run, graded in the open
We test many ways of finding earthquake risk. Each one gets its own record, here, with its misses. Most do not beat the normal odds yet, and we say so. A method joins the public watch only when its own record beats its own normal odds by a clear margin, checked forward in time. A high hit rate alone never qualifies.
The public 7-day watch
Our headline. The areas rising above their own normal, posted daily, scored against the normal odds.
On the record, graded every week
The models behind the record
Each model below logs its forecasts before the outcome. Its hit rate is not the test: in most of the places and sizes these models cover, a quake comes most weeks anyway. The test is whether the chances it gives beat just quoting the usual rate. We score that on forecasts the model was not tuned on.
- Loading the records...
Sealed tests
Written down first, graded later, no score until then
For each of these we wrote the test and what counts as a pass before the data came in, and timestamped it. We do not peek, and we do not show a running score that could be read as a result. Each is graded on its date, pass or fail.
- Sealed
The 7-day watch, for quakes of M5.0 or bigger
Do the watch's areas get a quake of the stated size more often than their own normal odds, by at least half again? Watch posts from Oct 15, 2026 to Jan 13, 2027.
Graded on or after January 24, 2027.
- Sealed
Quake clustering against our older watches
On quakes of M5.0 or bigger that start a sequence, does the standard quake-clustering model catch more of them than our older watches, using the same amount of map? Window Oct 12, 2026 to Jan 10, 2027.
Graded on or after February 15, 2027.
- Sealed
A year-ahead map for M5.5 or bigger
Does blending quake clustering with 20 years of past quakes place the next year's M5.5+ quakes better than clustering alone? Issued once on Oct 26, 2026, for the year after.
Graded on or after November 24, 2027.
- Sealed
A stricter watch list
A shadow list that names an area only when its clustering rate is at least double its normal. Does it hit at three times its own normal odds? Kept out of public posts until graded. Runs Oct 20, 2026 to Oct 20, 2027.
Graded on or after October 30, 2027.
- Sealed
Will a bigger quake follow?
After every quake of M5.0 or bigger, a sealed chance that a larger one follows within 7 days, scored against the standard method agencies use. Live since Oct 13, 2026; a second version built on quake clustering joined on Oct 20, 2026.
Graded on or after October 13, 2028.
- Sealed
Aftershock forecasts beside the official ones
After every M6.0 or bigger quake, we seal our own aftershock forecast next to the U.S. Geological Survey's, and score both as the sequence plays out. Where the USGS issues none, in Latin America, a separate sealed test compares a general forecast with one that learns from the sequence.
No skill claim until many sequences are in. The Latin America test is graded at 20 sequences or on October 10, 2028, whichever comes first.
- Sealed
Does a rise in risk come before quakes?
Whether an area's risk score rises more in the 72 hours before an isolated quake than at matched quiet times. Locked Aug 8, 2026, for quakes after that date.
Graded once 120 days have passed and 300 cases are in, first possible on December 6, 2026.
- Sealed
Two checks on our own aftershock and swarm models
Do our swarm detector's size ranges beat ranges of the same width from quake clustering? And do our aftershock models fall short of the standard clustering model they are built on? Both locked in August 2026.
The swarm check is graded on November 12, 2026, or once 100 forecasts are in, whichever is later. The aftershock check is graded once 120 days and 200 cases are in, first possible on December 6, 2026.
Retired, retracted or on hold
What did not hold up
These stay on the record. A list of methods that only shows the ones that worked would tell you nothing.
- Retired
How high a model's risk score is
When we checked our CRITICAL alerts in October 2026, they hit about as often as luck would. The level of the score carries no warning. What carries information is a score RISING above its own normal, which is what the public watch uses.
- Retired
Thousands of Earth signals on top of quake clustering
Adding our many sensor signals to the standard quake-clustering model did not improve it in any test we ran: by number of signals, by place, or by size. The two models that rest on them stop writing on October 25, 2026, and their records stay.
- Retired
Three stress meters from small-quake patterns
We tested three popular signs of stress in the crust, read from the sizes, clustering and rhythm of small quakes. In Oklahoma, where the stress from deep fluid injection is known, none tracked that stress beyond the plain quake rate, and none warned of M4 quakes. Retired on October 19, 2026.
- Closed
A deep-earth stress meter after great quakes
A model of how the deep earth slowly relaxes after a great quake, tested on Japan's 2011 quake. It did not fit the ground motion near the fault, and far away it could not be told apart from other effects. Closed on October 19, 2026.
- On hold
Changes in seismic wave speed
Small changes in how fast seismic waves travel through the ground. An audit found problems in how the measurements were processed, so the method is suspended and we publish no results from it until the audit is done.
- Retracted
The physical warning sign we sealed in 2026
It did not survive our own checks. The signal turned out to be the earthquakes recording themselves. Read the retraction →
- Retracted
A space-weather signal before slow slip in Cascadia
An early result linked a space-weather index to slow fault slip in the Pacific Northwest. We retracted that result. Its sealed forward test stays locked and is graded as written on May 22, 2028.
Methods still in private testing are not listed until their own tests are done.