cosmicraysandearthquakes/results/prereg_predictions.md

Pre-Registered Predictions — Out-of-Sample CR–Seismic Validation

Written: 2026-04-23T22:56:22Z Git SHA: unknown OOS window: 2020-01-01 → 2025-04-29 Surrogates: 100,000 phase-randomisation

This file was created BEFORE loading or analysing any out-of-sample data. All thresholds are pre-specified. Results are recorded in results/out_of_sample_report.md.

---

In-sample context (1976–2019)

From scripts 02–05 (Homola replication + stress tests):

Quantity	Value
Dominant peak lag (raw)	−525 days (half solar cycle)
Dominant peak |r| (raw)	0.469
r(τ=+15 d) raw	+0.310 (solar-cycle confounded)
r(τ=+15 d) HP-detrended	+0.041
In-sample p_global (IAAFT, raw)	1.000 (NOT significant after surrogate correction)
After detrending	p < 0.001 at lags ≠ +15 d

The in-sample dominant peak is at −525 days, not at the claimed +15 days. r(+15 d) ≈ 0.04 after solar-cycle removal — this is the baseline expectation for the out-of-sample window.

---

Pre-registered predictions

P1 — Sign and location of claimed correlation peak

Prediction: If Homola et al.'s mechanism is real, the OOS window should show a cross-correlation peak at τ ≈ +15 days (cosmic rays leading seismic activity by 15 days) with positive sign (positive CR deviation → elevated seismic Mw-sum 15 days later).

Operationalisation:

• PASS if r(τ=+15 d) > 0 AND the lag of maximum |r(τ)| for τ ∈ [5, 30] days is within ±3 days of +15 days.
• FAIL otherwise.

Baseline from in-sample HP-detrended: r(+15 d) ≈ +0.041 Monte Carlo tolerance (at 100,000 surrogates): ±0.0063

P2 — Significance and solar-phase trend

Prediction: The OOS window (2020–2025) covers Solar Cycle 25 rising phase, approaching the predicted 2025–2027 solar maximum. Homola's model predicts the CR–seismic correlation should be in a RISING phase of its ~11-year envelope (the last in-sample envelope peak was near 2014).

Operationalisation:

• PASS if: (a) p_global (phase-surrogate) < 0.05, AND (b) r(τ=+15 d) in rolling 18-month windows shows a non-negative trend (slope ≥ 0) across the OOS period.
• PARTIAL if (a) holds but (b) does not.
• FAIL if p_global ≥ 0.05.

P3 — Rolling-window lag stability

Prediction: The lag at which r(τ) is maximised for τ ∈ [5, 30] days should be stable to within ±3 days across 18-month rolling windows of the OOS data.

Operationalisation:

• PASS if std(τ*) ≤ 5 days across rolling sub-windows where a peak in [5, 30] days exists.
• FAIL if std(τ*) > 10 days or peaks migrate outside [5, 30] days in majority of windows.

P4 — Geographic non-localisation

Prediction: Per Homola et al.'s own result, the correlation should be GLOBAL (disappear in location-specific analyses). After BH FDR correction at q=0.05, the number of significant (station, cell) pairs should NOT significantly exceed the expected false-discovery count.

Operationalisation:

• PASS if n_significant ≤ 2 × expected_FP (BH q=0.05).
• FAIL if n_significant > 2 × expected_FP AND a clear geographic cluster emerges.

---

Falsification criteria (pre-specified)

F1 — No peak in claimed window

Criterion: No lag τ ∈ [5, 30] days has |r(τ)| exceeding the 95th percentile of the phase-surrogate distribution.

• F1 TRIGGERED (Homola falsified) if the criterion holds across the full OOS window AND across all 18-month sub-windows.

F2 — Peak lag drift

Criterion: The optimal lag τ* for τ ∈ [5, 30] days drifts by more than ±10 days between any two adjacent 18-month rolling windows.

• F2 TRIGGERED if drift > 10 days in majority of window pairs.

F3 — Unexpected geographic localisation

Criterion: The OOS correlation is STRONGER in a specific geographic region than globally — the inverse of Homola's own finding.

• F3 TRIGGERED if n_significant > 3 × expected_FP AND a geographic cluster with min p < BH-threshold is identified.
• This would be informative negative evidence: a real local effect, but NOT the global cosmic-ray mechanism Homola proposed.

---

Analysis decisions (pre-specified)

Parameter	Value	Reason
Bin size	5 days	Matches Homola et al.
Lag range	±200 days	Covers claimed +15 d with context; shorter window makes ±1000 d infeasible
Surrogates	100,000	GPU-accelerated; MC tolerance ±0.0063
Surrogate method	Phase randomisation	Preserves power spectrum; faster than IAAFT
Detrending	Linear + sunspot OLS	HP/STL inappropriate for <1 solar cycle window
Min stations/bin	3	Matches Homola et al.
Min magnitude	4.0	Matches Homola et al.
Rolling window	18 months	Minimum for meaningful correlation at 5-day bins
Rolling step	3 months	Smooth time evolution
FDR	BH q=0.05	Standard

---

This file is part of a pre-registered analysis. Results are reported regardless of direction in results/out_of_sample_report.md.