CLIMASENSE

AETHER

CLIMATE INTELLIGENCE

Autonomous Engineering Technology, Health and Environmental Risk

From environmental observations to accountable financial models.
Global, location-based climate intelligence for environmental exposure research and asset integrity

AETHER · HURRICANE GUIDANCE

Storm Tracking and Forecast Evaluation

Monitor storm activity, compare AETHER guidance with NHC forecasts, and review historical model performance.

Experimental research
Automatic checks while this page is open

Checks every 15 minutes while visible. Unattended capture is shown as active only after a scheduled run reaches the archive.

FORECASTS PRESERVED—Original capture and model retained
VERIFIED FORECAST PAIRS—Exact-time NHC preliminary observations
LATEST CHECK · UTCAwaiting first captureNo capture recorded

STORM ARCHIVE

Automatic Storm Archive

NHC-listed Atlantic, eastern Pacific and central Pacific systems are discovered by ID, with no fixed storm list or season cutoff. Every available matching forecast cycle is saved when a check runs.

Missing or stale inputs remain visible as gaps. An empty list does not indicate zero development risk.

Official NHC outlooks
Latest capture coverage and source gaps

Awaiting a completed NHC check.

Checks capture the latest available advisory, rather than every intermediate revision. Scheduled and page checks can fail; the archive records those gaps. The view lists the latest 200 snapshots.

PROSPECTIVE MACHINE LEARNING

Forecast Evaluation and Model Learning

Separate ridge regressions learn east–west and north–south baseline errors at 24, 48 and 72 hours. Learning begins from forecasts preserved here; the 2025 replay is not counted as new live evidence.

HorizonPairs / stormsBaseline errorNHC errorCandidate errorCandidate skill vs NHC
24 h0 / 0Awaiting outcomes—Collecting training pairsNot established
48 h0 / 0Awaiting outcomes—Collecting training pairsNot established
72 h0 / 0Awaiting outcomes—Collecting training pairsNot established
Learning method and evaluation rules

β = (XᵀX + 10I)⁻¹Xᵀy, separately for the east and north residual. The six inputs are an intercept, capped east/north differences between NHC and the baseline, normalized initial latitude, and sine/cosine of initial longitude. These are inputs available when the forecast is captured.

Each horizon needs 20 eligible training pairs before its candidate is shown; this is an engineering threshold, not evidence of statistical adequacy. The most recent 5,000 verified pairs per horizon form a bounded rolling training window. No target after the forecast cycle enters its model. Adjustments are capped at 2 km per forecast hour. Every snapshot saves its training IDs, coefficients and version.

Only exact-time tropical or subtropical preliminary BEST observations score a forecast. A record captured after its valid time, an NHC fallback, or a missing observation cannot train the learner. Predictions are scored before their outcome enters later training. Ended storms are checked for 14 days after their last saved forecast. Later source revisions do not rewrite an earlier score or model.

Candidate skill compares only the same cases with NHC. Repeated forecasts from one storm are dependent; preliminary means do not demonstrate a reliable improvement. The candidate remains experimental and never silently replaces the baseline model. This feature does not predict storm genesis or intensity.

2025 Historical Benchmark and Methodology

2025 ATLANTIC · CHRONOLOGICAL HOLDOUT

Current method: the original six-storm benchmark

20% improvement target: not met

Methods were selected on earlier storms, then evaluated on later storms. Positive skill means lower mean error than NHC on the same cases. These are research-sample results, not NHC’s basin-wide published averages.

Mean great-circle position errors and uncertainty in the average skill
HorizonNHCGDMI + fallbackAETHERSkill 95% intervalNHC fallback cases

All three AETHER skill intervals include zero. The correction worsened the underlying GDMI result at 24 and 48 hours; at 72 hours AETHER retains GDMI unchanged. The observed advantage over NHC cannot be attributed solely to AETHER.

Method, data and limits of interpretation

Forecast Method

The chronological forecast method uses NHC-hosted GDMI as its starting position. At 24 and 48 hours it applies 25% of a time-weighted recent forecast bias, calculated against the current storm analysis. At 72 hours the GDMI position is unchanged. Without GDMI, the same-cycle NHC position is displayed as an explicit fallback.

The weighted bias uses forecasts made 12, 24 and 48 hours earlier, valid at the current cycle. The adjustment is capped at 2 km per forecast hour. Future observed positions are used only for scoring the historical replay.

b(h) = Σ[e^(−l/24) min(h/l, 2) d(l)] / Σ[e^(−l/24)]
Δ(h) = cap(0.25 b(h), 2h km), for h = 24 or 48
Δ(72) = 0
Skill = 100 × (1 − mean AETHER error / mean NHC error)

Here l is an available lag in hours and d(l) is the east/north displacement from the older GDMI forecast to the current analysis. A spherical destination calculation applies Δ. No available bias terms means zero adjustment.

Interpretation and Limitations

Training cutoff: 1 September 2025. Historical evaluation: 290 lead-specific comparisons from 120 cycles across six storms, with five storms qualifying at 72 hours. Repeated cycles are not independent storms. The 95% intervals resample whole storms, with 2,000 paired bootstrap samples and fixed predictions.

Archive receipt times were not verified. Historical reconstruction is not evidence of a live 2025 deployment, and 2025 results do not establish 2026 performance. The model does not predict storm formation, category or landfall probability.

GDMI is the NHC-postprocessed WeatherNext Cyclones guidance used in the 2025 archive. AETHER’s correction is a separate calculation. This view does not run Aurora or coupled ocean–atmosphere fluid dynamics.

Solar entropy, Bonamy Grades from the Bonamy Matrix, and warmest-water traces are not inputs to this tested track model. The ocean observations remain a separate dated view; no historical SST snapshot is presented as contemporaneous forecast evidence.

Exported records retain their cycle, valid time, model attribution and source. Current NHC availability is checked when the site opens, after a stale tab becomes active again, and on request; stale, conflicting or mismatched inputs do not generate an AETHER track.