NUFORC UFO Reports Don't Cluster Around US Earthquake Epicenters in a 104-Year Public-Record Cross-Match

Dek. We tested the population-level prediction of the "earthquake lights" folklore by cross-matching 9,379 US M ≥ 4.5 earthquakes (1910–2014; effective volume concentrated 1995–2014) against 71,008 NUFORC UFO sighting reports; observed coincidences within 100 km and -24 h / +1 h of epicenters (3) sit well below the random-shuffle null mean (9.4) and the null finding survives every sensitivity check.

Workspace slug: earthquake-lights-nuforc-usgs

Status: v2 — Sasha review applied (one CRITICAL + three IMPORTANT catches addressed).

Date: 2026-05-30.

---

Abstract

The folklore claim known as "earthquake lights" — visible luminous phenomena (flashes, glows, ball lightning, columns of light) appearing in the hours before or during significant earthquakes — has been recorded in seismology literature since at least the 1965 Niigata event and continues to be cited in popular media. If the claim holds at population scale and at the unstratified level of an open eyewitness-reporting network, NUFORC reports of sky anomalies should cluster in space and time around earthquake epicenters at a rate above chance.

We cross-matched the US M ≥ 4.5 earthquake catalog (ANSS ComCat, 1910–2014, n = 9,379 after deduplication) against the US-resident NUFORC UFO sighting database (n = 71,008 in joint window) using a 100 km spatial radius and a -24 h / +1 h temporal window around each earthquake epicenter. We then permuted UFO timestamps 10,000 times, preserving the spatial distribution of reports (and therefore the population-density and reporting-bias structure of the null), and recomputed the coincidence count for each permutation.

Across 458,960 spatial pairs (UFO report within 100 km of an epicenter), we observed 3 coincidences in the folklore-relevant time window. The permutation null mean was 9.44 ± 3.51 (95% CI [3.0, 17.0]). The rate-ratio is 0.318 (95% CI [0.176, 1.000]); p = 0.989 in the folklore-supporting direction. The largest events in the catalog (M ≥ 6, n = 756) produced 0 coincidences against an expected 0.11. No magnitude bin between M 4.5 and M 9.2 shows enrichment, and the null finding survives a sensitivity sweep across window width (-48 h to -6 h), magnitude floor (M ≥ 3.5 to M ≥ 5.5), and the restricted 1995-2014 sub-window that carries the bulk of NUFORC's reporting volume.

We conclude that the earthquake lights folklore, in the specific form "US M ≥ 4.5 earthquakes produce sky anomalies that the NUFORC submission population would notice and submit at a population-level rate above chance", is not supported by the public record. The test is statistically powered to rule out large effect sizes (rate-ratio ≳ 1.9 at 80% power); it does not address the geology-specific predictions of the leading proposed mechanism (Theis et al. 2014), it does not address sub-NUFORC-threshold luminous phenomena, and it does not preclude a real folklore signal that is masked by an earthquake-induced depression of NUFORC submission rates.

Introduction

Earthquake lights are an old folklore claim with a small but persistent scientific literature. Theis et al. (2014) catalogued 65 historical accounts back to 1600, and proposed a piezoelectric mechanism linking subbasalt stress to atmospheric ionization. The mechanism remains contested and the empirical base — historical eyewitness accounts at named events — has not, to our knowledge, been tested against a population-level cross-match.

The TerraPulse platform stores both layers needed for such a test:

• NUFORC (National UFO Reporting Center): 80,332 self-reported sighting records, 1906–2014, of which 71,040 are US-located with lat/lon coordinates.
• USGS Earthquake Catalog: globally complete from the early 20th century via the ANSS ComCat web service.

If the folklore signal is real at the magnitude that would produce population-level reports, NUFORC submissions should be enriched within a small spatial radius of and a few hours before significant earthquakes, relative to a null that preserves the spatial and temporal distribution of UFO reports under random matching.

Data

NUFORC

National UFO Reporting Center reports, ingested into TerraPulse from the public NUFORC archive. Each record carries an alleged-sighting timestamp_utc, latitude, longitude, US state, and a free-text description. The submission timestamp is stored separately in the record metadata and is not used in this analysis.

Coverage in PG: 1906-11-10 to 2014-05-08, n = 80,332 globally. US-only with non-null coordinates: n = 71,040, distributed across all 50 states plus DC.

NUFORC volume is heavily concentrated post-1995, reflecting both the growth of the NUFORC submission system (online forms from the late 1990s) and the overall population-level uptake of UFO reporting. Pre-1990 volume is sparse; pre-1970 is anecdotal.

USGS earthquake catalog

ANSS ComCat (Advanced National Seismic System Comprehensive Catalog), queried via the FDSN web service for events between 1906-11-10 and 2014-05-08, magnitude ≥ 4.5, within a bounding box that covers the United States (latitude 18° to 72° N, longitude −180° to −65° E). Events were post-filtered to US territory via the catalog's place field, requiring a match against the list of US state and territory names.

We applied the TerraPulse canonical earthquake deduplication: round timestamp_utc to the nearest minute, group by (time, latitude, longitude, magnitude), keep first. This removes ComCat's documented duplicate records arising from overlapping product feeds.

Final catalog: n = 9,379 US M ≥ 4.5 events, 1906-11-15 to 2014-05-05.

Joint window

The intersection of both layers is 1910-01-01 to 2014-05-05, a span of 104.3 years. We restrict the analysis to this window. After filtering, n_quakes = 9,379, n_ufo = 71,008.

Method

Spatial-temporal cross-match

For each earthquake k, let N_k be the set of NUFORC reports within 100 km of the epicenter (great-circle distance). For each report u ∈ N_k, the pair (k, u) is a coincidence if

t_q − 24 h ≤ ts(u) ≤ t_q + 1 h

where t_q is the earthquake time and ts(u) is the alleged-sighting time of the report. The window is asymmetric (24 h before, 1 h after) because the folklore predicts pre-event luminosity; we cap the after-window at 1 h to capture syn-event reports while excluding rubbernecking and news-cycle bias.

Spatial neighbours are computed using a BallTree with the haversine metric on (latitude, longitude) in radians. We enumerate all 458,960 spatial pairs as a flat table for the permutation step.

Permutation null

The null preserves the spatial distribution of UFO reports — which encodes population density, urban concentration, and reporting-network bias — by holding each report's latitude and longitude fixed. We shuffle the UFO timestamp vector globally and recompute the total coincidence count. We repeat this for 10,000 permutations using a numpy RNG seeded at 20260530.

Under H0, the expected total coincidence count is small because the 25 h window is a tiny fraction of the 104 y span; the null distribution we observe (mean 9.44, std 3.51) reflects this.

Test statistic

We report:

1. Rate-ratio. observed / mean(permuted), with a 95% CI from the 2.5th–97.5th percentiles of the permutation distribution.
2. p-value. Fraction of permutations with coincidence count ≥ observed, plus one to avoid zero (Phipson and Smyth 2010).

We also report per-magnitude-bin rate-ratios for M bins [4.5, 5.0), [5.0, 5.5), [5.5, 6.0), and [6.0, ∞) to test for the folklore-predicted magnitude dependence.

Results

Headline

• Observed coincidences across the 104 y joint window: 3.
• Permutation null mean: 9.44 ± 3.51 (95% CI [3.0, 17.0]).
• Rate-ratio: 0.318 (95% CI [0.176, 1.000]).
• p = 0.989 in the folklore-supporting direction.

The observed count sits in the lower tail of the permutation distribution. A two-sided depletion test would put the inverse direction at p ≈ 0.011 (marginal), but the absolute count is small enough that we do not interpret this as evidence of avoidance; we treat the finding as a clean null on the folklore-supporting alternative.

Magnitude dependence

The folklore mechanism predicts a stronger signal at higher magnitude. We observe the opposite pattern: the M ≥ 6 catalog (756 events, several of which were felt by millions and extensively photographed) shows zero coincident NUFORC reports against an expected 0.11. The single bin with a rate-ratio above 1 (M 5.5 – 6.0) carries 2 observed against 1.50 expected, which is not significant (p = 0.45).

Concentration

The 3 observed coincidences are: 2 in California (both 1999) and 1 in Virginia (2011). Geographic and temporal concentration of the observed count is meaningless at n = 3, but for completeness: California carries ~14% of US NUFORC volume in the window and ~37% of US M ≥ 4.5 events; its 67% share of the observed coincidences is consistent with chance.

Sensitivity analyses

We repeated the test across five parameter variants to check whether the headline null is window- or floor-specific. The result is robust:

Observations:

• The null finding does not depend on the 25 h window choice. At ±48 h (wider) the observed count rises to 9 but the null rises proportionally to 19.7; the rate-ratio is essentially flat at 0.46. At ±6 h (tighter) the observed count is 0 against a null of 2.9.
• The null finding does not depend on the M ≥ 4.5 floor. Dropping the floor to M ≥ 3.5 quadruples both observed (95) and expected (124) into a regime where the test has its tightest measurement; the rate-ratio remains below 1 (0.77, p = 0.977).
• The M ≥ 5.5 sub-bin shows a rate-ratio of 1.24 (2 observed against 1.62 expected), the only variant where the point estimate sits above 1.0. This is not statistically significant (p = 0.48) and rests on n_obs = 2.
• The 1995-2014 restriction confirms the headline finding survives on the bulk-volume sub-window. The pre-1995 sparse era is not driving the null.

Discussion

What this rules out

The test rules out a folklore signal of the magnitude that would produce a clean population-scale rate-ratio above ~1.8 against the null. Our detectable effect floor is set by the small absolute null count (9.44), which itself is determined by the 25 h time window applied across a 104 y span. Within those limits, the public record does not support the claim that visible luminous phenomena precede or accompany US M ≥ 4.5 earthquakes at a rate detectable by the NUFORC reporting network.

The M ≥ 6 bin merits particular attention. The folklore is most often invoked for "great" earthquakes — events large enough to crack lithospheric stress to depths where the proposed mechanisms (piezoelectric, p-hole) could operate at observable amplitudes. Our catalog includes 756 such events over the joint window, including the 1964 Anchorage (M 9.2), 1989 Loma Prieta (M 6.9), and 2002 Denali (M 7.9) earthquakes. NUFORC carries 0 coincident reports within 100 km and 24 h prior of these events; the expected count under random matching is 0.11. The folklore mechanism predicts more, not fewer, reports here.

What this does not rule out

1. The Theis-mechanism prediction is not directly tested. The leading mechanistic account (Theis et al. 2014) ties earthquake lights to specific geological substrate (p-hole-rich subbasalt, rift environments) and fault geometry (vertical-stress events: thrust, normal). The Theis catalogue's historical lights concentrate in the European rift, Mississippi Embayment, and Italian Apennines. Our M-bin sweep is unstratified by geology or fault type; a real signal confined to those mechanism-favorable subsets would not be detectable in our test. A flat null on a flat M-bin sweep is consistent with "no signal anywhere" and with "signal only in geology-specific subsets we did not separately analyze." This test discriminates between those two interpretations only weakly.
2. Earthquake-induced NUFORC submission depletion. The most parsimonious explanation for the observed depletion direction (rate-ratio 0.318, observed in the lower tail of the null) is not a real avoidance phenomenon but a confound: significant earthquakes cause local power outages, internet outages, and observer distraction. People in or near the disaster zone do not submit to NUFORC in the hours after a quake, and may be too occupied to submit in the hours before if foreshock disturbances disrupt routines. This mechanism would deplete the post-event window directly and the pre-event window partially. The permutation null preserves spatial structure but does not preserve this earthquake-conditional submission-rate disruption; the disruption would survive into the null only if it operated identically on permuted dates, which it does not. We treat this as the load-bearing alternative for the depletion-direction signal: it is consistent with the data, it does not require positing a real avoidance phenomenon, and the test as designed cannot separate it from a "true zero."
3. Small effect sizes in the folklore-supporting direction. Rate-ratios between roughly 0.8 and 1.4 are compatible with the data given the small absolute counts. A genuine folklore signal at, say, a 20% enrichment would not be cleanly detectable in this test. (See power discussion below.)
4. Non-NUFORC observer populations. NUFORC reports are filtered through self-identification as a UFO observer. The set of people who see a bright light during an earthquake and report it to NUFORC may differ systematically from the set who see one and report it to local news or to a seismology institution.
5. Sub-NUFORC-threshold phenomena. The folklore literature includes accounts of brief, faint, and diffuse glows that would not provoke a NUFORC submission. Our test addresses what NUFORC's submission population reports, not what every observer perceives.
6. Sub-hour temporal precision. The mainstream folklore is "minutes before to during" the quake. Our 25 h window is generous in both directions. A tighter window (e.g., ±15 min) might detect a phenomenon that our window dilutes; but with our absolute counts a tighter window would also reduce power.

Power

Under the empirical null mean of 9.44 and std of 3.51, the minimum detectable rate-ratio at 80% power (one-sided, α = 0.05) is approximately (9.44 + 2.49 × 3.51) / 9.44 ≈ 1.93. At 50% power the floor drops to ≈ 1.61. The test is therefore well-powered to detect a large folklore signal and poorly-powered to detect anything below a ~60% enrichment. The four-bin magnitude sweep adds five tests to the implicit family (1 headline + 4 bins); Bonferroni-corrected, none of the bin p-values approach significance in either direction. We report uncorrected p-values throughout for transparency.

Caveats

• NUFORC volume non-uniformity. Submission volume is concentrated in 1995–2014. The vast majority of our 458,960 spatial-pair "opportunities" come from this sub-window.
• Diurnal mismatch. NUFORC reports cluster at night; earthquake times are uniform across the day. Our 25 h window covers ≥ 1 full nightcycle per quake, so we expect this to average out, but the day-of-week distribution of NUFORC (weekend skew) interacts with the uniform-by-week quake distribution in a way that is not fully decoupled.
• NUFORC timestamp precision. Sighting times are minute-precise for some submissions and rounded to the hour or "evening" for others. The 25 h window is robust to this; a tighter follow-up would not be.
• US-only. The result is specific to NUFORC's US-language reporting culture. Other UFO-reporting populations (UK BUFORA, French GEIPAN) might show a different pattern.

Comparison to prior TerraPulse work

This is the second TerraPulse cross-match paper to find a clean null on a UFO-related folklore claim. The Skinwalker Ranch cross-match paper (workspaces/skinwalker-ranch-baseline/cross-match) found that 7 public sensor layers stay quiet across 10 filming-day pins of The Secret of Skinwalker Ranch. Both findings cut against the same family of claims: that anomalous-witness reports correlate with underlying geophysical events.

The two papers differ in scope and statistical posture:

• The Skinwalker paper tests 84 specific witness claims against multi-modal sensor coverage; it concluded 0 TESTED + 2 WX-CONTEXT + 24 UNTESTABLE + 58 NO-PIN.
• The earthquake lights paper tests a single mechanism (UFO ↔ quake coincidence) against a large and well-indexed pair of public catalogs across 104 y.

Both find that the public record does not corroborate the underlying anomaly claim.

Conclusion

We tested one specific population-level prediction of the earthquake lights folklore: that significant US earthquakes should produce eyewitness sky-anomaly reports clustered in space and time around their epicenters at a rate above chance, as reflected in the NUFORC submission record. They do not. Across 104 years of US M ≥ 4.5 earthquakes (n = 9,379) and the US-resident NUFORC sighting record (n = 71,008), observed coincidences (3) sit below the permutation null mean (9.44); no magnitude bin between M 4.5 and M 9.2 shows enrichment; the M ≥ 6 sub-catalog produces zero coincident NUFORC reports against an expected ≈ 0.11; and the result is robust across a sensitivity sweep of window width, magnitude floor, and the 1995-2014 bulk-volume sub-window.

This negative result is bounded. It rules out a large-effect-size folklore signal in the NUFORC submission population, at the population-aggregated scale, without geological stratification. It does not test the Theis et al. (2014) mechanism's geology-specific predictions; it does not address sub-NUFORC-threshold luminous phenomena; it does not address non-NUFORC observer populations; and it does not exclude an earthquake-induced depression of NUFORC submission rates that would mimic an avoidance signal regardless of whether earthquake lights are real. The proposed physical mechanism for earthquake lights is therefore not ruled out by this test — only its predicted footprint in this particular eyewitness-network record, and only at the magnitude floor where the test has power.

Reproducibility

• Scripts: scripts/extract_nuforc.py, scripts/extract_quakes.py, scripts/analyze.py.
• Data outputs: data/ufo_us.parquet, data/quakes_us_m45.parquet, data/coincidences_observed.parquet, data/permutation_null.parquet, data/results.json.
• ComCat fetch is per-year cached at data/comcat_cache/{year}.json for re-runs.
• RNG seed: 20260530.

References

1. Theis, F., et al. (2014). Prevalence of earthquake lights associated with rift environments. Seismological Research Letters 85 (1): 159–178.
2. Phipson, B., and Smyth, G. K. (2010). Permutation P-values should never be zero. Statistical Applications in Genetics and Molecular Biology 9 (1).
3. ANSS Comprehensive Catalog. U.S. Geological Survey. https://earthquake.usgs.gov/data/comcat/
4. National UFO Reporting Center. https://www.nuforc.org/