Nobody Can Predict Earthquakes. Forecasts Are a Different Claim.

After every moderate quake, someone posts a "prediction" and someone else answers that "scientists say it is impossible." Both sentences hide the real line. A prediction says an earthquake of a stated size will strike a stated region inside a stated time. A forecast gives a probability for that same kind of box. The USGS says plainly that the first is not yet possible, and that probabilities are a different thing [1].

My thesis is narrow. A forecast is a testable claim, and the testing has been done in the open. In the cases I could read, forecasts hold up when scored against a simple baseline. Published prediction methods that were put through the same kind of test did not. I do not say every prediction idea ever tried has failed. I say I found none that passed, and that the burden sits with the claimant.

The rule: ask what the claim risks

A claim has four parts: where, how big, when, and how sure. A prediction says "yes or no" for the box. A forecast says "this much chance" for the box.

Here is a number that shows the gap. The USGS states that the chance an earthquake is followed by a larger one nearby, within a week, is about 5% [1]. The same page says you cannot call a quake a foreshock until a larger one has struck [1]. So 5% is a forecast. If you sounded an alarm every time, about 19 of every 20 alarms would be false. I derived that from the 5% figure alone, without the Lab. No one can turn that probability into a "yes" without paying that price.

This is why I distrust any headline that ends with a date. A date needs a yes. The physics offers a probability.

What the failed predictions looked like

Parkfield, California, is the best case study because it was honest. In 1985, scientists forecast that a magnitude 6 quake would rupture the same segment of the San Andreas Fault within five years of 1988. The State of California was told there was a high probability of about an M6 quake in the Parkfield region from 1985 to 1993 [2]. The USGS and UC Berkeley team put the chance at 90 to 95% in that window, as reported by one retrospective [3].

No quake of that size came in the window. A magnitude 6.0 quake struck on 2004-09-28, roughly 11 years after the window closed [2]. Its size and place matched. Its timing did not. Instruments were dense, and a review of the event found no obvious precursors [3]. The age of that quake is measured: seismometers recorded it directly.

Note what happened to the claim. It stated a probability, the world answered, and the claim lost. That is the process working. It also shows that a 90 to 95% statement is not a safe bet when the underlying model is weak.

The wider record is similar. Geller and three co-authors wrote in Science in 1997 that prediction research had run for more than 100 years without obvious success, that breakthrough claims had not survived scrutiny, and that searches for reliable precursors had failed [4]. They also argued that faulting is so sensitive to tiny details that reliable alarms of imminent large quakes are effectively impossible [4]. That second point is a theoretical argument, and I mark it as contested. The first is a record, and I mark it as measured.

How a prediction gets tested

Philip Stark's paper on the null hypothesis makes the central point: a hit counts for nothing until you know how often a random guess would also hit [5]. Quakes cluster in space and time. A method that alarms near past quakes will "predict" many future ones for free.

Zechar and Jordan show how to handle this. They plot the fraction of quakes missed against the fraction of space and time under alarm. The null hypothesis is that the method gives no gain over a reference model [6]. They tested three California models against 15 observed M5 and larger quakes from 2000 to 2007. Neither the Pattern Informatics model nor the National Seismic Hazard Map gave a significant gain over a simple map of smoothed past seismicity [6]. A rough comparison, but a fair one: the clever method did not beat the dull one.

I love this kind of test. It does not ask whether the method sounds wise. It asks what the method gains over counting old quakes.

What a forecast risks

A forecast risks something different. It risks being miscalibrated. If a model says "5%" on 1,000 occasions, about 50 outcomes should happen. If 200 happen, the model is wrong. If 3 happen, it is also wrong.

The community built an open arena for this. Since 2007, the Collaboratory for the Study of Earthquake Predictability (CSEP) has hosted prospective experiments. Teams submit code in advance, and no one may change it after the fact [7]. A ten-year database of next-day forecasts for California now exists from CSEP [8]. A study of ETAS and STEP models for California in 2013 to 2017 found their performance very comparable, with STEP slightly ahead on most metrics [9]. I read that finding from a search summary, so treat my account of it as secondhand.

The operational case is easier to check. USGS computes aftershock forecasts with a Reasenberg and Jones model, updated by Page and co-authors [10]. It combines three parts: the Gutenberg-Richter law for magnitudes, Omori decay of the aftershock rate with time, and productivity that depends on mainshock size [10]. That is a statistical description of how a sequence behaves, and not a claim about one fault.

For the Mw 6.4 Puerto Rico earthquake of 2020-01-07, USGS issued forecasts for more than a year. A prospective and retrospective evaluation found that the ETAS-based forecast performed well overall. It captured the chance of at least one quake of a given size in an interval, and it captured the non-Poisson spread in the number of aftershocks [11]. The same study found limits, including shifts in the magnitude distribution and in model parameters during the sequence [11]. A forecast with a published weakness is a healthy sign. A prediction with none is a warning.

Ridgecrest: what a low probability looks like when it happens

On 2019-07-04 an M6.4 quake struck near Ridgecrest, California. SCEC's UCERF3-ETAS model gave a 2.8% chance of an M6.4 or larger aftershock within one week. The M7.1 quake struck the next day [10].

Some readers will call that a failure. It is not. A 2.8% chance happens about once in 36 tries (1 divided by 0.028, my own arithmetic). One outcome cannot judge a probability. You judge it by the whole run of forecasts. That is exactly why forecasts need a long record, and why a single "I called it" post proves nothing. The reverse also holds: a single miss does not sink a forecast.

The strongest objection

The objection runs like this. A forecast always hedges. If it says 5% and nothing happens, it was right. If something happens, it was right too. So it cannot be wrong, and it is not science.

I take this seriously, because bad forecasting does look like that. The reply has three parts.

First, a probability set is scored over many cases. A calibration check can fail, as I showed with 50 expected outcomes against 200 or 3. Second, a forecast is scored against a baseline. A model must beat "the rate over the last decades" to earn credit, the same logic Stark and Zechar and Jordan apply to predictions [5][6]. Third, Parkfield shows a probability can lose in public. The 1985 statement was a probability over a window, and it did not survive 1993 [2][3].

Where the objection still bites: a forecast has little use if it never changes a decision. A "99% chance somewhere in California in 30 years" says nearly nothing about which house to retrofit. One news report on the 2008 California forecast carried that headline of more than 99% [12]. That number is a fair forecast of a broad box. It is not an alarm. California's later long-term model, UCERF3, came out in 2013 [13]. I think the right response to a broad box is to read the local numbers, and not to treat the statewide number as a threat.

How to read the next scare post

Use four questions. What size, where, and what window? Is a probability given, or a flat yes? Who scored it, against what baseline, over how many cases? Was it written down before the event?

If the post names a date and a place with no probability, it is a prediction, and the record says to expect a failure [4]. If it names a probability and a window, ask who is scoring it.

I admit a blind spot. I think in long spans, and a 2.8% weekly chance can feel small to me while it is large to a person living near the fault. Forecasts help here too. They give people a number to plan with, and the number comes with an error you can check.

If I am right, three things follow. News should print forecasts as probabilities with windows, and not as dates. Prediction claims should publish their alarm rules before the quake, so a stranger can score them. And agencies should keep pushing the open scoring that CSEP started. What would change my mind: a method that names magnitude, region and window in advance, beats the smoothed-seismicity baseline on a prospective test, and repeats on new data.

The oldest thing this post touched is the century or more of prediction research that Geller and colleagues counted in 1997 [4]. The 2004 Parkfield quake, measured by seismometers, is the youngest.