The Earthquake-Lights
Signal That Isn't

The Folklore

Earthquake lights are an old folklore claim: visible flashes, glows, ball lightning, and columns of light reported in the hours before or during significant earthquakes. The claim is recorded in seismology literature since the 1965 Niigata earthquake in Japan, and continues to surface after every notable quake in popular media. A 2014 paper by Theis et al. proposed a piezoelectric mechanism, in which stress on certain rock types generates electrical charge that ionizes air. The mechanism remains contested.

If the claim holds at population scale, meaning real people see real things during real earthquakes, then the simplest population-level signature is that eyewitness reports of weird sky phenomena should cluster around earthquake epicenters. The TerraPulse platform stores both layers needed to test that prediction. This is a story about how that test comes out positive — and why the positive is not real.

The Test

We cross-matched two public catalogs, both read from the TerraPulse dexes through the platform provenance registry:

• USGS (U.S. Geological Survey) earthquakes, magnitude (M) ≥ 4.5, located in US territory, 1910–2014. After deduplication: 9,379 events, from the eq dex.
• NUFORC (National UFO Reporting Center) sightings, US-resident, with coordinates, 1910–2014: 70,854 reports, from the nuforc dex. Each carries an alleged sighting time, a latitude and longitude, and a state.

For each earthquake, we counted NUFORC reports within 100 km of the epicenter and within the 25-hour window from 24 hours before to 1 hour after the quake time. The asymmetric window reflects the folklore claim that earthquake lights appear before or during the quake. To know whether a count is "a lot," we shuffle: keep every report's location fixed (preserving population-density bias) and randomly re-assign its timestamp, 10,000 times, recomputing the count each time. That random distribution is the null.

The Apparent Signal

The count is sharply positive. Across the 104-year window we observe 38 coincidences against a random-shuffle mean of 13.0 — a rate-ratio of 2.9 (p = 0.0002). And it gets stronger with magnitude, which is exactly what the folklore mechanism predicts:

• M 4.5–5.0: 24 observed vs 9.3 expected — 2.6×
• M 5.0–5.5: 7 observed vs 2.5 expected — 2.8×
• M 5.5–6.0: 4 observed vs 0.8 expected — 5.0×
• M ≥ 6.0: 3 observed vs 0.4 expected — 7.3×

Taken at face value, this is a population-scale confirmation of earthquake lights: significantly more sky-anomaly reports near epicenters, strongest at the great earthquakes the folklore most often invokes. If we stopped here, we would publish "earthquake lights confirmed." So why is the headline of this piece the opposite?

Trap One: Counting the Same Report Seven Times

The statistic counts pairs — every (earthquake, report) match. But big earthquakes come with dense aftershock sequences, all clustered at the same spot. A single UFO report near the mainshock gets counted again against every aftershock. One report, seven quakes, seven "coincidences."

The 38 coincidence pairs resolve to just 14 distinct UFO reports. When we count each report once — a report near a mainshock and its six aftershocks counts once, not seven — the result is 14 observed vs 11.1 expected, a rate-ratio of 1.26, p = 0.22. Not significant. The moment you stop double-counting, the signal is gone.

Trap Two: It's Two Earthquakes

Where does even the inflated count come from? Two places. Of the 38 pairs, 33 are in California, and by year, 1989 and 1994 account for 26 of them. Those are Loma Prieta (October 1989, M 6.9, San Francisco Bay Area) and Northridge (January 1994, M 6.7, Los Angeles) — two of the most consequential, most densely-populated, most heavily-reported earthquakes in US history.

Remove those two sequences (30 quakes) and recompute:

• Pair count, full catalog: 38 vs 13.0 → rate-ratio 2.93
• Pair count, without the two mega-quakes: 12 vs 12.0 → rate-ratio 1.00
• Distinct reports, full catalog: 14 vs 11.1 → rate-ratio 1.26
• Distinct reports, without the two mega-quakes: 11 vs 10.9 → rate-ratio 1.01

The entire "signal" is two earthquakes. A test powered by two events is not a population-scale test; it is two anecdotes with a p-value. And both events came with wall-to-wall media coverage, power outages, and transformer arcs — Loma Prieta struck during a live World Series broadcast. A report of a strange light in San Jose that evening is as consistent with a rattled city looking up as with a geophysical glow.

Sensitivity

The apparent positive is not a lucky window or magnitude cutoff — it persists across the parameter sweep, because California genuinely hosts many earthquakes and many UFO reports. That robustness is exactly why the confound controls above, not the parameter sweep, are the analysis that matters.

• Wider window (−48 h to +1 h): 63 vs 25.6 — 2.5×
• Tighter window (−6 h to +1 h): 19 vs 3.6 — 5.3×
• Lower floor (M ≥ 3.5): 310 vs 136.8 — 2.3×
• Higher floor (M ≥ 5.5): 7 vs 1.2 — 5.8×
• 1995–2014 only: 8 vs 9.1 — 0.9×, no signal

The last row is the tell. The 1995–2014 sub-window is the only variant that excludes both Loma Prieta (1989) and Northridge (1994) — and it is the only one with no signal. Cut out the two mega-quakes by any means, whether by name or just by date, and the "signal" is gone.

What This Does and Doesn't Say

The confound-controlled result is a clean null: counting each report once, and setting aside two mega-quakes, US earthquakes produce no excess of sky-anomaly reports. But it is a bounded finding. It does not rule out:

1. The Theis mechanism's geology-specific predictions. The leading account ties earthquake lights to specific subbasalt geologies and fault types. Our test is unstratified by geology; a signal confined to mechanism-favorable subsets would not show up in a flat magnitude sweep.
2. Faint, sub-threshold phenomena. Some accounts describe brief, dim glows a witness would notice but never report to a UFO database. We measure what NUFORC's submission population reports, not what every observer perceives.
3. Any individual event. This is a statement about the population-scale rate, not a claim that nothing luminous ever happened at a specific earthquake — Loma Prieta in particular has a genuine eyewitness-lights literature. It says only that such events are too few and too confounded to establish a rate above chance.

The real lesson is methodological. A naive spatiotemporal cross-match handed us a significant positive with a textbook magnitude curve. It survived a five-way parameter sweep. And it was entirely an artifact of double-counting and two famous quakes. The honest test is the confound-controlled one — not the first number that clears p < 0.05.

Reproducibility

All scripts, extracted data, the permutation null distributions, and the results JSON are in the TerraPulse repository at workspaces/earthquake-lights-nuforc-usgs/. Both layers are read from the TerraPulse dexes (nuforc, eq) through the provenance registry. The naive pair count is in analyze.py; the distinct-report and mega-quake-exclusion controls are in reanalyze_confound.py. The random seed (20260530) is recorded in the scripts.

The full LaTeX paper, with the null-distribution figure, the magnitude-bin table, and the confound decomposition, is at paper/paper.pdf in the workspace.