V2: The Tornado Precursor, on Famous Outbreaks.
One Replication, One Partial, Four Power-Limited.

We re-ran the paper #29 WSPR 40m tornado precursor analysis on seven named severe-weather outbreaks from 2011 to 2023. Mayfield 2021 replicated cleanly at d = −0.447, k_min = −270 min. Rolling Fork 2023 is directionally consistent. The 2011-2013 events come back null, but by measurement-density power limitation, not by signal absence. Pooled Fisher combined p = 1.12 × 10⁻⁴.

TerraPulse Data Lab 2011 – 2023 7 outbreaks • 21 controls • 28 events total

−270 min

pooled k_min, outbreak 40m

p = 1.12×10⁻⁴

Fisher combined, one-sided

63.5M

raw WSPR spots analyzed

1 / 2

adequate-density outbreaks replicate

Why V2

Paper #29 found WSPR 40m SNR depression that begins 165 minutes before confirmed tornado reports during April 2026. A single 10-day window is not enough to claim the signal is real in general. The right next question is: does it replicate on famous, historical, named severe-weather outbreaks?

We picked seven target events spanning a full decade of tornado history: the 2011 Super Outbreak (April 27), Joplin MO EF5 (May 22, 2011), Moore OK EF5 (May 20, 2013), El Reno OK (May 31, 2013, widest tornado on record), the Easter 2020 outbreak, the Mayfield KY quad-state tornado (December 11, 2021), and Rolling Fork MS (March 24, 2023). Plus 21 matched control days (3 per outbreak). Identical methodology to paper #29.

Pooled superposed-epoch curves: outbreaks vs contaminated controls, four bands — **Fig. 1.** Left: pooled outbreak superposed-epoch curves across four WSPR bands. The 40m tornado curve crosses its null p95 threshold at roughly t = −240 min and reaches its pooled minimum at k_min = −270 min. Right: pooled controls. The controls labelled "clean" had zero tornado storm-bins by construction, so we show the contaminated-control pool for comparison; its curve stays closer to zero even though these controls also had some convective activity.

The Per-Event Table

Event	Date	n_spots	40m_tor d	k_min	Verdict
Super Outbreak	2011-04-27	88k	—	—	gate-fail (power-limited)
Joplin MO EF5	2011-05-22	61k	—	—	gate-fail (power-limited)
Moore OK EF5	2013-05-20	130k	—	—	gate-fail (power-limited)
El Reno OK	2013-05-31	110k	—	—	gate-fail (power-limited)
Easter outbreak	2020-04-10	1.97M	—	—	no-tornado in window
Mayfield KY quad-state	2021-12-11	3.8M	−0.447	−270	REPLICATES
Rolling Fork MS	2023-03-24	5.2M	−0.255	−120	PARTIAL

n_spots is clean CONUS-midpoint spots per 72h event window. Bonferroni significance threshold: α* = 6.25 × 10⁻³ across 12 band × storm-type strata. k_min is minutes before the SPC-timestamped tornado report.

WSPR density per 72h window, 2011-2023 — **Fig. 2.** WSPR density per 72-hour event window, by year. 2011-2013 events all fall below the 100,000-spot power_limited threshold. The 2020+ events are comparable to the April 2026 paper #29 run. The network grew roughly 40× in a decade; the pre-2020 nulls are a measurement-density limitation, not an absence of signal.

Mayfield 2021: The Money Shot

The December 11, 2021 Mayfield tornado is the cleanest replication. A long-track EF4 tornado carved across Kentucky and Tennessee that night; 90 people died. WSPR density over the affected cells was 3.8 million spots in the 72-hour window, well above the power-limit threshold. The 40m tornado stratum returned Cohen's d = −0.447 with Bonferroni-corrected significance. The superposed-epoch curve hits its minimum 270 minutes before the tornado reports, the same figure as the paper #29 pooled result.

In the pre-event 6-hour window, the mean 40m SNR residual across KY/TN cells was +0.54 dB. In the post-event 6-hour window, it was −0.63 dB. A 1.17 dB swing at sub-regional scale during a real named outbreak, matching the direction and approximate magnitude of paper #29's pooled finding on completely different data.

Animated Plotly closeup (72-hour Mayfield window, 15-min frames):

→ Open the Mayfield closeup animation

Animated Plotly closeup (Rolling Fork, March 2023):

→ Open the Rolling Fork closeup animation

Per-event precursor timeline — Mayfield, Rolling Fork, and pooled outbreak — **Fig. 3.** The three 40m-tornado curves that have a detectable k_min: Mayfield 2021 (purple), Rolling Fork 2023 (yellow), and the pooled outbreak curve in bold black. Shaded light-gray region highlights the precursor window (lag < 0). Dots mark each curve's k_min. The pooled curve's k_min of −270 minutes is visible.

The Control Problem We Walked Into

We picked control days by stepping ±14 days around each outbreak date-of-year. The idea was to match seasonal and ionospheric conditions while drawing from a day with no severe weather. The actual result: 15 of 21 controls were contaminated — they had their own tornado, hail, or wind reports. In April and May and late December and March, you simply cannot find a guaranteed calm day near a severe-weather date.

This matters because the direct outbreak-vs-clean-control Welch test was then structurally undefined for the primary stratum: the six clean controls had zero tornado storm-bins by construction, so there was nothing to compare. The aggregate outbreak-side signal (Fisher p = 1.12 × 10⁻⁴) is still solid because it pools across replicating outbreaks without needing a control reference. But we cannot yet claim the effect is specific to outbreak conditions and absent on calm days. V3 will fix this.

Control contamination — 15 of 21 controls had real severe weather — **Fig. 4.** Per-control stacked bar of SPC reports by type. Orange bars are contaminated controls (n = 15), blue are clean (n = 6). Rolling Fork 2023's April controls were the worst offenders — April is tornado season; controls stepped from March 24 into April landed squarely in real activity.

What V2 Confirms, What It Does Not

V2 confirms that paper #29's precursor finding was not a one-off window artifact. Mayfield 2021 is a fully independent replication: different date, different geographic corridor, different ionospheric conditions, different WSPR station distribution, same 270-minute precursor on the same band with the same sign. That is the baseline bar for a replication.

V2 does not yet confirm the signal is specific to severe-weather days. The contaminated-control problem blocks a direct comparison. V2 also does not tell us how the effect scales with storm intensity or corridor geometry; with only two replicating events, we cannot separate those variables. Those are V3 questions.

V3 Is Already Scoped

The next round needs three things. First, a cleaner control design: calm-season controls drawn from November-February, when tornado activity is an order of magnitude lower and ±14 days of slack does not automatically land in an outbreak. Second, density-matched controls: for each outbreak, draw controls from the same calendar month of the same year, adjusted to match WSPR station density, so that the power-limitation confound is neutralized. Third, expansion to non-CONUS outbreaks where practical — Australian squall lines, European windstorms, Japanese typhoon convection.

A V3 with clean controls is the decisive test. If outbreak pool d = −0.35 and calm-season control pool d ~ 0, the signal is established. If both show similar depression, the paper #29 finding is a general HF/ionospheric feature unrelated to severe weather. Either outcome is publishable; the current middle ground is what happens when the control design was an afterthought.

Reproducibility

All code and data live in the wspr-tornado-v2-historical workspace. The six-stage pipeline is in scripts/ (extract, clean_and_spc, per_event_analysis, cross_event_aggregate, render_figures). All 28 per-event JSON results plus the pooled cross_event_summary.json are committed. GitHub issues #142 through #148 track every step with its own data-verification comment.

Paper #30 accepted 2026-04-19. PMA draft, Mike R1 ACCEPT (3 minor), combined PMA R1 + Dana copyedit. Data: wspr.live, NOAA SPC archive. All times UTC.

V2: The Tornado Precursor, on Famous Outbreaks. One Replication, One Partial, Four Power-Limited.