Summary
1 · The starting point, corrected
Genesis PI is informative — because of geography
What a genesis descriptor says about the eventual peak
Mann–Whitney AUC, all tropical cyclones (major vs not), 95% season-bootstrap interval. ERA5 at the exact genesis hour, 200–800 km annulus.
2 · The persistence trap
Most later "skill" was distance to the threshold
What 48–72 hours of intensity history actually adds
Brier-score improvement on five-season holdouts. Grey: adding current wind (distance to 96 kt). Orange: adding the preceding 24-h change, beyond current wind. Bars: 95% intervals.
3 · Emanuel (2000), revisited globally
Storms reach a uniformly random share of their ceiling
Fraction of storms whose normalized peak ν = vmax/vp exceeds x
Hurricanes not limited by declining PI within 72 h of peak. Monthly climatological ERA5 PI (Emanuel's choice). A uniform distribution is a straight line (dashed).
The ceiling and the peak rose together
Season means for not-limited hurricanes: PI at the place and month of peak (actual month) and the peak wind itself, m s⁻¹. Lines: least-squares trends.
4 · Two clocks
A race between the core and the runway
The longer the runway, the likelier a major
Fraction reaching 96 kt by runway length (days from genesis to landfall or to water with PI < 96 kt). Storms whose archive ended first (censored) are excluded here.
The race, storm by storm
Hurricanes with an environmental end to the runway (random sample of 900). Above the diagonal the runway outlasts the time it took to reach hurricane strength.
Why did the peak stop?
What followed each storm's first lifetime-maximum fix, by peak class.
5 · What the vortex remembers
Alignment matters most; size barely at all
Parent-circulation descriptors at genesis
AUC for reaching major (left) and RI onset within 72 h of genesis (right). Hollow: raw. Filled: ranked only against storms forming at similar latitude.
6 · Adding it up
From genesis, a little; with the road ahead, a lot
Held-out AUC for eventual major intensity, as information is added
Five-season holdouts (solid) and whole-basin holdouts (hatched where shown). The last bar uses the observed future runway — an upper bound for what a track forecast could add, not a genesis predictor.
7 · Follow-ups
A real forecast recovers part of the runway; the rest of the checks change nothing
Held-out AUC for eventual major intensity, 2017–2025, at the first forecast after genesis
Context + PI at the forecast time, then + the runway along the forecast track (to 120 h), then + the observed runway capped at the same 120 h. Five-season holdouts.
8 · Why near-misses stop on open water
Thin warm water and rising shear — not the ceiling
Near-misses and majors, 24 hours either side of the same wind speed
Medians. Near-misses (orange): hurricanes that peaked at 64–95 kt with at least a day of major-capable ocean ahead, centered on their peak. Majors (blue): each crossing a wind drawn from the near-misses' peaks, centered on that crossing. Shear and SST: ERA5 at the exact hour, 200–800 km ring. TCHP: ORAS5 monthly at the center.
What separates stopping from continuing, at the same intensity
AUC of each descriptor at the anchor time, oriented so 0.5 is no separation; 95% season-bootstrap interval. Right: near-miss vs major medians, and which group is higher. Gray: intensity · violet: atmosphere · teal: ocean · brown: structure.
Where the shear comes from
Near-misses move into the shear; it does not arrive over them
How much more the 850–200 hPa shear rose for near-misses than for future majors at the same wind (m s⁻¹, mean difference), split exactly into the change along the track at fixed times (moved into) and the change in place at fixed positions (arrived). ERA5 200–800 km ring winds read at all nine combinations of the storm's position and the time, 24 h either side of the anchor.
The daily check
Daily ocean against monthly: what each adds, out of sample
Change in held-out AUC when each set of ocean descriptors is added (dot; bar = 95% season-bootstrap interval). Filled: five-season holdouts. Hollow: whole-basin holdouts. Monthly: ORAS5. Daily: GLORYS12, sampled 3 days before the storm's day. Storms since 1993.
9 · Does a storm's peak scale with its ceiling?
Unit scaling is an assumption — but the strongest storms come close
Elasticity of lifetime-maximum intensity to potential intensity, by the competing-hazard ratio r
d ln LMI / d ln PI for the median (blue) and 90th-percentile (orange) peak of a storm seeded at the observed median 0.17 of its PI. Solid: the seed is a fixed fraction of PI. Dashed: the seed is a fixed wind. Shading: the range of r estimated from 1979–2025 storms; line: the estimate from the shape of the peak distribution.
How strong is the external clock? r by basin
Filled: r fitted to the shape of each basin's normalized-peak distribution (maximum likelihood with each storm's own seed; bar = 95% profile interval). Hollow: r from the runway clock that ends only at land or 45°, divided by the median intensification rate.
Which part of the ceiling's history matters?
How far back along the track intensity remembers potential intensity
Left: out-of-sample R² of the 12-hour wind change on the reduced-model drive PI² − V², with PI taken where the storm was L hours earlier (right of zero) or will be (left of zero, beyond the 12-h step a placebo); bars: ridge weights of a joint model over the past 72 h. Right: out-of-sample R² of ln(peak wind) on PI averaged over the storm's life before the peak with weight e−age/τ, against single-time choices. Five-season block cross-validation.
Emanuel's model on the real tracks
How fast the peak rises with the ceiling in FAST, run along every observed track
Elasticity of the population's peak wind to a 10% rise in potential intensity, with shear, dryness and track held fixed. Filled: 99th-percentile peak (bar = 95% season-bootstrap interval); hollow: median peak. Shaded: the two-clock model fitted to the record (§9 table, 99th percentile); dashed: Emanuel's (2007) synthetic storms, top decile.
10 · Open questions
Method
References
- Emanuel, K. (2000). A statistical analysis of tropical cyclone intensity. Mon. Wea. Rev. 128, 1139–1152.
- Kaplan, J. & DeMaria, M. (2003). Large-scale characteristics of rapidly intensifying tropical cyclones in the North Atlantic basin. Wea. Forecasting 18, 1093–1108.
- Gilford, D. M. (2021). pyPI v1.3. GMD 14, 2351–2369.
- Leipper, D. F. & Volgenau, D. (1972). Hurricane heat potential of the Gulf of Mexico. J. Phys. Oceanogr. 2, 218–224.
- Zuo, H., Balmaseda, M. A., Tietsche, S., Mogensen, K. & Mayer, M. (2019). The ECMWF operational ensemble reanalysis–analysis system for ocean and sea ice (ORAS5). Ocean Sci. 15, 779–808.
- Lellouche, J.-M., et al. (2021). The Copernicus global 1/12° oceanic and sea ice GLORYS12 reanalysis. Front. Earth Sci. 9, 698876.
- Emanuel, K. (2007). Comment on "Sea-surface temperatures and tropical cyclones in the Atlantic basin" by P. J. Michaels, P. C. Knappenberger and R. E. Davis. Geophys. Res. Lett. 34, doi:10.1029/2006GL026942.
- Emanuel, K. (2017). A fast intensity simulator for tropical cyclone risk analysis. Nat. Hazards 88, 779–796.
- Lin, J., Rousseau-Rizzi, R., Lee, C.-Y. & Sobel, A. (2023). An open-source, physics-based, tropical cyclone downscaling model with intensity-dependent steering. J. Adv. Model. Earth Syst. 15, e2023MS003686; code at github.com/linjonathan/tropical_cyclone_risk.
- Sparks, N. & Toumi, R. (2024). The Imperial College Storm Model (IRIS) dataset. Sci. Data 11, 424.
- Hersbach, H., et al. (2020). The ERA5 global reanalysis. QJRMS 146, 1999–2049; read from the Google ARCO-ERA5 store.
- The companion paper: What's Special About Major Hurricanes.