A week of earthquakes, and a record that keeps changing

Everything on this page is computed in your browser from three public USGS feeds, with no server in between. This section covers the feeds, how they are merged into one catalog, the statistics behind each chart, and where those statistics stop being reliable.

Feeds, merging and revisions

The page loads the past week once (about 1.4 MB, roughly 2,000 events), then polls the past hour feed every 60 seconds and the past day feed every 5 minutes. All three are GeoJSON summary feeds from the USGS Earthquake Hazards Program. Every response is merged into one catalog keyed by event id, and anything older than seven days is dropped.

Each event carries an updated timestamp. When a poll returns a larger one, the page compares the two versions and logs which of magnitude, magnitude type, review status, depth, epicenter (moves of 0.5 km or more), place name, or event type changed. Updates that touch none of these, such as new felt reports, are counted but not listed. An equal or smaller updated is ignored, because the feeds are cached separately and an older copy can arrive after a newer one.

When another network's solution becomes preferred, the event's id changes. The feed lists every associated id, so the page logs this as a revision of the same event (“id A to B”) rather than a deletion plus a new event. An event that was in the previous past day response, is missing from the current one under all of its ids, and is still at least 10 minutes inside the 24 hour window is logged as deleted. If a response built after that comes back with the event, it is restored and the log says so. A past day response that is not newer than the previous one is never compared, since it would lack the events added in between.

Explosions, quarry blasts and other non tectonic events stay in the merge and the revision log, but are left out of every count, chart and fit.

Magnitude of completeness

Small earthquakes are missed more often than large ones, so below some magnitude the catalog is incomplete and the frequency magnitude distribution bends over. Every estimate here uses only events at or above the magnitude of completeness, Mc. Mc is found by maximum curvature (MAXC): the 0.1 magnitude bin with the most events. MAXC tends to underestimate Mc, so the page adds 0.2, the correction recommended by Woessner and Wiemer (2005). MAXC is attempted only with at least 50 earthquakes in the region.

b-value

Above Mc, magnitudes follow the Gutenberg-Richter law. The b-value is the Aki (1965) maximum likelihood estimate with Utsu's correction for magnitudes rounded into bins of width dM = 0.1; its standard error is the Shi and Bolt (1982) formula.

The bootstrap resamples the whole regional catalog with replacement 200 times, recomputes Mc on every resample, and refits b. The interval is the 2.5th to 97.5th percentile of those values, with a fixed seed so the same catalog always gives the same interval. Because Mc is recomputed each time, the interval carries the uncertainty in Mc, which the Shi and Bolt error does not. A b-value needs at least 50 events at or above Mc; resamples that fall short are dropped and counted.

The test suite checks this on synthetic catalogs with a known b = 1. Over 300 catalogs of 5,000 events the estimates averaged 0.9965 with a spread of 0.0183, close to the mean Shi and Bolt error of 0.0176, and the bootstrap interval covered the true value 93.7% of the time. A single seeded catalog can still land outside its own interval, so the suite checks coverage across 20 catalogs instead of trusting one.

Aftershocks and the Omori-Utsu law

For every earthquake of M 4 or more in the region, the page draws the space and time window of Gardner and Knopoff (1974) and counts the later events inside it. A candidate followed by a larger event inside its own window is a foreshock and is skipped. The mainshock with the most aftershocks is the one plotted.

For M 4 the time window is already 41 days, longer than the catalog, so in practice every later event within distance d counts. With 30 or more aftershocks, the rate n(t) = K / (t + c)^p is fitted by maximum likelihood (Ogata 1983) on the aftershock times t_i, in days since the mainshock, over [0, T], where T runs from the mainshock to now. K is profiled out, which leaves a search over c and p.

c is searched on a logarithmic grid of 61 values from 0.001 to 1 day and p from 0.40 to 2.00 in steps of 0.02. The footer gives the 95% profile likelihood range of p (where 2 (log L_max - log L) ≤ 3.84) and says when the best value sits on the edge of the grid. The dots are counts in logarithmic time bins, four per decade, divided by the bin width. Empty bins cannot be drawn on a log axis; the table view lists them.

With few aftershocks p is loose. On synthetic sequences with p = 1.1 and c = 0.05 days, p varied by 0.21 across catalogs of 120 aftershocks over 3 days and by 0.065 at 1,000; the profile range covered the true p in 95% and 93% of 200 catalogs. The suite therefore checks that the range covers the true value rather than holding the point estimate to a fixed tolerance.

Busy hours

Each UTC hour's count k is tested against a Poisson baseline. The rate lambda is the median of the previous 24 hourly counts, so the burst being tested does not inflate its own baseline. In quiet regions that median is often zero, so the mean of the same 24 hours is used instead, and lambda is never below 1/24 (one event a day). An hour is flagged when P(X ≥ k) < 0.001 under Poisson(lambda). The first 24 hours of the week have no full baseline and are not tested; the current hour is partial and is tested as it stands. A flag marks clustering, usually an aftershock sequence or a swarm. It is not a forecast.

Limitations

• The feed is not one catalog. US regional networks report down to M 0 or below with local (ml) and duration (md) magnitudes, while elsewhere only events of roughly M 4 and above appear, mostly as body wave (mb) and moment (mww) magnitudes. Mc and b for All events mostly describe that mixture, so a departure from b = 1 there says more about the catalog than about the crust. Even inside one region, ml and md are not the same scale.
• One week is short. Regional b-values from a few hundred events carry wide intervals, and a single sequence can dominate a week.
• Right after a large earthquake, small events are hidden in the mainshock's coda and its early aftershocks, so completeness is temporarily worse than the single weekly Mc suggests. In the Omori-Utsu fit this early deficit is absorbed into c, and all magnitudes in the window are used.
• Automatic solutions change. Magnitudes move by tenths of a unit, types change, and some events are deleted, so numbers computed in the first minutes after an event are provisional. The revision log shows this directly.
• The Gardner-Knopoff windows were calibrated on Southern California; elsewhere they are a convention. Region boxes are rectangles in latitude and longitude, and coastlines are Natural Earth at 1:110 million, coarse at regional zoom.