It looks from Figure 3 that they fit a sinusoidal curve to the data, and then fit an exponential to the residuals of *that* fit. That exponential shows a sharp increase from hour -2, but the curve is not a good fit to the rest of the data, and it almost has to show an increase somewhere because of the functional form.

I would have assumed the way to evaluate this method would be to back-test it: figure out when in the past this would have predicted that there would be an earthquake, and measure the accuracy/precision/etc. of that prediction.

It often surprises me how often disciplines don't have take a predictive approach by convention: if you make predictions at least it's possible to be *wrong*.

Yikes this travesty of a website with all the ads and popups makes this page impossible to read.

Link to the source: https://www.science.org/doi/10.1126/science.adg2565

Omg that web site is horrendous. Had to give up reading the article.