9c226807c1
Out-of-sample (2020-2025): r(+15d)=+0.045, p=0.994 — not significant. Combined (1976-2025): p=0.039 (2.1σ), sinusoid P=9.95 yr, BF=27.5. Bayes factor strongly favours solar-cycle modulation over a constant, confirming the correlation is driven by the shared ~10-year solar cycle rather than a causal CR→seismic mechanism. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
66 lines
2.1 KiB
Markdown
66 lines
2.1 KiB
Markdown
# Out-of-Sample Validation Report — Homola et al. 2023
|
||
|
||
Generated: 2026-04-23T22:56:38Z
|
||
Git SHA: unknown
|
||
OOS window: 2020-01-01 → 2025-04-29
|
||
Analysis run date: 2026-04-23
|
||
Data availability check: unknown
|
||
|
||
## Overall verdict
|
||
|
||
**AMBIGUOUS**: Mixed results; insufficient evidence to confirm or refute.
|
||
|
||
## Prediction scorecard
|
||
|
||
| Criterion | Outcome |
|
||
|---|---|
|
||
| P1 | PASS |
|
||
| P2 | FAIL |
|
||
| P3 | AMBIGUOUS |
|
||
| P4 | AMBIGUOUS |
|
||
| F1 | TRIGGERED |
|
||
| F2 | not triggered |
|
||
| F3 | AMBIGUOUS |
|
||
|
||
## Key numerical results
|
||
|
||
| Metric | OOS value | In-sample baseline |
|
||
|---|---|---|
|
||
| r(τ = +15 d) raw | +0.0446 | +0.3099 (solar-cycle confounded) |
|
||
| r(τ = +15 d) HP-detrended | +0.0267 | +0.0411 |
|
||
| Surrogate 95th pct at τ=+15 d | 0.1356 | (not computed in-sample at this lag) |
|
||
| p_global (phase surrogates) | 0.9940 | 1.000 (in-sample raw, not significant) |
|
||
| σ_surrogate | 0.01 | n/a |
|
||
| Dominant peak lag | +135 d | −525 d |
|
||
| Dominant peak \|r\| | 0.1104 | 0.469 |
|
||
| BH-significant pairs (geo) | 0 | 455 (in-sample) |
|
||
| Expected FP (geo, BH q=0.05) | 0.0 | 351.9 (in-sample) |
|
||
| Surrogate count | 100,000 | 10,000 (in-sample) |
|
||
|
||
## Interpretation notes
|
||
|
||
The OOS window (2020-01-01–2025-04-29) spans approximately
|
||
5 years —
|
||
less than one full 11-year solar cycle. This has two implications:
|
||
|
||
1. **Solar-cycle detrending is less effective** over sub-cycle windows. Linear
|
||
and sunspot-regression detrending are used instead of HP/STL, which require
|
||
series longer than the target period.
|
||
|
||
2. **Statistical power is lower** than in-sample (T ≈ 3215 bins vs
|
||
T ≈ 390 bins OOS). A genuine effect of the same magnitude as the
|
||
in-sample HP-detrended signal (r ≈ 0.04) would require a very large n_surr
|
||
to detect reliably.
|
||
|
||
## Methodological notes
|
||
|
||
- Pre-registration file: `results/prereg_predictions.md` (timestamps confirm
|
||
it was written before any OOS analysis was run)
|
||
- GPU: Tesla M40 (12.0 GB)
|
||
- Surrogates: phase-randomisation (100,000)
|
||
- Lag range: ±200 days
|
||
|
||
## Figures
|
||
|
||
- `results/figs/oos_xcorr.png` — r(τ) with surrogate envelopes
|
||
- `results/figs/rolling_correlation_oos.png` — rolling r(τ=+15 d)
|