The Number of Storms

Section 1

The question nobody has answered

Write the annual global count as N. Three families of explanation exist for its magnitude and stability: supply theories (the tropics seed a certain number of precursor disturbances; Emanuel 2022, Hsieh et al. 2022–2024), geometry theories (rotating radiative-convective equilibrium fills a domain with cyclones at a characteristic spacing; Khairoutdinov & Emanuel 2013, Chavas & Reed 2019), and recovery theories (each storm consumes its patch of ocean-atmosphere for a time; Vu et al. 2021, Zhou & Lin 2024). They make different, measurable predictions about the statistics of the count, the spacing of simultaneous storms, and the aftermath of each storm's passage. The best track can adjudicate all three. That is what this page does.

Section 2

A metronome, measured: the count is as steady as randomness allows

The stability of N is celebrated, but stability against what benchmark? For a count of independent rare events at a fixed rate — a Poisson process — the variance equals the mean, so a rate of ~ per year comes with a free standard deviation of ~. The proper metric is the Fano factor, variance divided by mean:

The observed F = (detrended ) sits below one — mildly sub-Poisson, though not individually significant (chance of a value this low under pure Poisson: ). Hurricane-strength storms give F = , major-strength F = . The year-to-year autocorrelation of the global count is — no memory. Elsner & Jagger (2013) noticed the same regularity in the 1981–2006 counts and guessed that feedbacks were responsible; Section 8 measures one.

The sharp statement hides in what F ≤ 1 forbids. If the planetary genesis rate itself wandered between years — warm decades making more storms, cold fewer — the count would be a mixed Poisson process, and mixing always adds variance: F = 1 + CV²·E[N], where CV is the interannual coefficient of variation of the rate. Simulating that process against the observed F:

The TC-year splices hemispheres (NH calendar + SH Jul–Jun), so as a robustness check the count was rebuilt by genesis date in common calendar windows for all twelve starting months: F spans — at or below Poisson in every alignment, in none of them significantly below.

Section 3

Basins trade; the globe nets

How can basin rates swing while the global rate holds? Bookkeeping: the variance of a sum is the sum of the variances plus twice the covariances,

The cancellation has a face: the North Atlantic anticorrelates with the East Pacific (r = ), the West Pacific () and the South Pacific () — partly the ENSO seesaw, but only partly: after detrending and regressing out the ONI, the summed cross-basin covariance falls from to storms², and the Atlantic–West Pacific anticorrelation survives untouched (ENSO barely moves the West Pacific count: El Niño , La Niña ). A permutation test (shuffling each basin's years independently, which preserves every basin's own statistics and kills only the synchrony) puts the observed global variance at the percentile of the independent-basin null ( after detrending) — suggestive rather than conclusive at 46 years, and Frank & Young (2007) found no significant cancellation in 1985–2003. The rate bound of Section 2, not this ledger, is the robust statement.

Section 4

The count is the stiff variable; the energy is loose

Global ACE and power dissipation each factor exactly into (number of storms) × (energy per storm). On log-variances, the shares of interannual variance are:

The count, with its CV of , contributes barely a sixth of the variance of global ACE. The rest is per-storm energy — lifetime (CV ) times intensity along the track — the quantities that float on where storms form, how long they live in high-potential-intensity air, and whether ENSO has parked genesis east or west. Storm-days per year (mean , equivalent to storms active on an average day) vary nearly twice as much as the count that produces them.

Section 5

What does not set the number, part I: the warm pool

The oldest intuition is thermal: storms need ≥26.5°C water, so the area of such water should set the population. That hypothesis has now run a 46-year natural experiment:

The ≥28°C pool grew by % over the record; the count trend (%/decade) is statistically flat, and detrended warm-pool area and count are uncorrelated (r = ). This falsifies proportionality to warm-pool area; it is a natural experiment, not a controlled one, and subtler thermodynamic influence is not excluded. The genesis threshold is relative to the tropical mean, not absolute — as the whole tropics warm, the bar rises with the water (Vecchi & Soden's relative-SST argument, here as a direct falsification of area control).

Section 6

Storms keep their distance: a measured exclusion zone

Condensed-matter physics asks of any collection of particles: given one at the origin, what is the chance of finding another at distance r, relative to random placement? That is the pair correlation function g(r), and it distinguishes a gas (g = 1 everywhere) from a liquid (a suppressed core out to the particles' contact distance). Here the particles are every pair of simultaneously active tropical cyclones in the same basin, 1980–2025. The "random placement" expectation counts, exhaustively, every same-basin pair whose two storms come from different seasons at the same day of year and hour — identical seasonal geography, no possibility of interaction — scaled so the ratio is absolute: g = 1 means exactly the pair count independent placement predicts.

Integrating the deficit over the core gives the area a storm denies to other storms' centers, expressed as a hard disk:

The uniformity is the surprise: six basins with different sizes, monsoon regimes and storm climatologies agree to within ±10%, the East Pacific — the basin of smallest storms — at the low end. The rotating-RCE literature (Khairoutdinov & Emanuel 2013; Chavas & Reed 2019) predicts cyclone spacing scales with a deformation radius ~1/f, shrinking steeply toward the poles. Binning the pairs by latitude:

Where the shell points: bearing and lag

An isotropic annulus is a weak signature of wave seeding, so resolve it by direction. For every genesis with an active same-basin storm at that instant, take the bearing from the existing storm to the new one in a hemisphere-mirrored frame (0° = poleward, 90° = east; the SH flipped so both hemispheres composite), against the same cross-season expectation:

Section 7

What does not set the number, part II: a shallow scar

Consumption models (Vu et al. 2021; Zhou & Lin 2024) propose that each storm leaves its patch dried, stabilized and cooled, and that the patch needs two to three weeks to recover — so the tropics can sustain only so many storms per season; Zhou & Lin's packing scaling reaches ~100 storms a year with a 2–3-week recovery. The recovery is measurable. For every genesis, count the earlier 6-hourly fixes of other storms within 500 km, by lag, against the same count from other seasons at the same day of the season; then divide by the same ratio for fixes 1,500–3,000 km away, which cancels basin-wide clustering (an active MJO phase lifts both rings):

The thermodynamic memory, twenty years on

Hart, Maue & Watson (2007) quantified a storm's local memory as the evolution of the maximum-potential-intensity anomaly from 20 days before passage to 60 days after, against an evolving climatology — weekly SST and coarse reanalysis were the best available. The same measurement on the daily 0.25° ERA5/tcpyPI store ( NH track points, Jul–Oct 1996–2025):

Section 8

The thermostat: occupied basins make fewer storms

An exclusion zone could merely rearrange genesis in space without touching the total rate — and the Rossby-wave-dispersion literature even has storms sowing successors 2,000–3,000 km downstream. The direct test: does a basin's genesis rate fall when the basin already holds storms? Every basin-day 1980–2025, with the expected rate taken from that basin's own day-of-year climatology:

Section 9

Slack at the peak

How close does the real atmosphere come to the exclusion-zone capacity? At the Northern Hemisphere's peak (Aug 15 – Sep 30), the hemisphere averages simultaneous storms, hits once a season, and has never in 46 years exceeded :

Two facts. First: the hemisphere never comes close to filling its favorable area — storms at most, against a capacity of ~ even under the most conservative closure. Hoogewind et al. (2020) found the same order-of-magnitude gap by packing storms at an assumed separation; here the separation is measured. Second: storms are not crowded even where they form. The median distance from a peak-season storm to its nearest same-basin neighbor is  km, exclusion distances. Against the same number of storms placed independently (positions drawn from other seasons at the same calendar moment):

Section 10

What does not set the number, part III: the ceiling nobody touches

Could energy supply limit the count? Define the thermodynamic ceiling as the power dissipation the tropics would deliver if every point of favorable ocean ran at its potential intensity — the same integrand the storms themselves obey, ρC_DV³, integrated over the favorable area:

Section 11

What sets the number, then

Assemble the pieces. The annual count factors as

with an areal bound n(t) ≤ n_max(t) = A_fav(t)/πr_eff². Run the arithmetic with measured values: were the Northern Hemisphere packed to capacity (~ storms at once, the most conservative closure) for its 120-day season, the hemisphere alone would produce ~ storms a year. The observed 88 for the whole globe sits at under a sixth of that bound. The count is therefore not set by any global limit — not area, not energy — and that is precisely what the statistics of Sections 2 and 4 already said: a hard-limited quantity would be pinned at its ceiling and over-dispersed below it, while the observed count is a fixed-rate Poisson stream, mildly under-dispersed by local interaction. The number is rate-set, locally regulated: seed production in the genesis belts fires at a planetary rate that holds to within a few percent across 46 years (the true constant needing explanation), and the ~1,150-km exclusion zones operate as a fast local thermostat wherever storms crowd — clipping bursts (Section 8), pruning close approaches (Section 9), keeping the Fano factor below one (Section 2) — with a shallow three-week scar behind each storm (Section 7), and without the globe ever approaching full.

So the answer to "why ~90?" decomposes into a bounded claim and an open one. The bounded claim, now with observational numbers attached: the number behaves as the product of a near-constant planetary seed supply and a survival gauntlet whose measured self-interactions — the exclusion zone, the occupancy thermostat, the shallow scar — are a regulation of where and when, not ultimately of how many per century. One identification caveat is owed: a steady count constrains only the product (seed rate) × (survival probability) — seed production could rise while survival falls — and the observational seed census of Ikehata & Satoh (2021) finds survival, not seed number, carrying much of the variability. The open claim is why that product is what it is — the standing problem of Sobel et al. (2021), which a seed census joined to a survival model, not a storm count, will have to close.

Could the best track supply that census itself, through the systems that became depressions but never reached gale strength? No — and the ledger shows why:

The global failed-storm record collapses from ~/yr to ~/yr while gale counts hold flat; the North Atlantic's falls by a factor of eight (naming and designation practice, not physics), and the Southern Hemisphere barely records failures at all. The best track censors its own denominator. A real seed census has to come from reanalysis, and the natural frame for it is the marsupial paradigm (Dunkerton, Montgomery & Wang 2009; the PREDICT campaign, Montgomery et al. 2012): a proto-storm develops inside the "pouch" of its parent wave — the critical-layer recirculation that travels with the wave and shields the embryo from shear and dry air. The pouch is exactly the object our measured fields act on: the exclusion zone and the wake are shear, subsidence and dry-air export — pouch killers — so the storm-attached suppression of Sections 6–8 should operate on pouch survival, not on wave production. That is the sharp, testable form of the supply-vs-survival question (Ikehata & Satoh 2021 already find survival carries most of the variability), and it extends to the baroclinic seed class — tropical transition and the diabatic Rossby vortex (Davis & Bosart 2003; Moore & Montgomery 2005), ~15–20% of geneses globally (McTaggart-Cowan et al. 2013) — whose "pouch" is a self-regenerating moist PV anomaly rather than a wave critical layer. The concrete next dataset: ERA5 wave-pouch tracks (Okubo–Weiss in the co-moving frame), each followed to genesis or death, joined to the occupancy, bearing and wake fields measured here.

What this page rules out is as useful as what it pins down: not warm-pool area, not energy supply, not areal packing of the favorable ocean — and environmental recovery only as a few-percent trim.

And the companion question — what controls global ACE and PD — inherits a clean answer: with N pinned, the energy metrics are per-storm thermodynamics. Where genesis happens relative to the high-PI reservoirs (the ENSO redistribution of Maue 2011 and of this site's power-dissipation climatology), how long storms live there, and how much of their own ceiling they use — with four orders of headroom above the whole enterprise. A theory of global ACE is a theory of placement and duration, not of number.

Section 12

Methods and caveats

Population: every storm in this site's season archive (JTWC + NHC + Neumann lineage best track) with at least one fix that is simultaneously of tropical or subtropical status and ≥34 kt, 1980–2025 — the conjunction is tested at the same fix, so a depression whose gale winds occur only while extratropical does not count; a storm-by-storm census ledger (each system's qualifying fix, and each excluded sub-gale system) ships beside this page as frequency_census.csv. Six-hourly synoptic fixes at ≥34-kt tropical intensity for all spatial statistics; SH seasons Jul–Jun labeled by ending year, with a twelve-phase calendar-window recount as the alignment check. The pair expectation counts every same-basin pair of storms from different seasons at the same day of year and hour, divided by (seasons − 1), so g(r) is absolute; it is then divided by its 3,000–6,000-km level to remove same-season clustering. r_eff integrates the deficit below 1,500 km into an equivalent hard disk (a center-to-center distance), with jackknife-by-season errors and a season bootstrap for the latitude exponent. The wake statistic is the same construction in space and time, with a far-ring (1,500–3,000 km) control for clustering; the obs/expectation ratio estimates the forward conditional genesis rate given a prior fix (a risk-set quantity, not a retrospective lag distribution), and the estimator was validated on a synthetic inhibition-only process (hard-core thinning at 1,150 km: core recovered, no spurious shell, no latitude trend). A sensitivity weighting each unique storm pair once moves r_eff by <1%. Genesis = first eligible fix. The occupancy test's expected rates come from each basin's own 21-day-smoothed genesis climatology, so seasonality cannot masquerade as feedback; its intervals come from a year-block bootstrap, and the vacated-basin control (days with no storm active, stratified by departures in the prior week) guards against intraseasonal-phase confounding. Favorable area and the dissipation ceiling use the daily 0.25° ERA5 potential-intensity store built for the power climatology (tcpyPI, CKCD = 0.9), Jul–Oct; realized dissipation uses the size layer's radii-constrained wind fields, so ceiling and realized share the ρC_DV³ normalization.

Caveats, honestly held: the sub-Poisson Fano factor and the covariance cancellation are suggestive (p ≈ 0.06–0.17) at 46 years, not conclusive — the bound on rate wander is the robust statement, and it is conditional on the assumed conditional-dispersion family (a sufficiently sub-Poisson conditional law can hide more rate variability). The North Atlantic's overdispersion is real and known (Conway–Maxwell–Poisson fits reach the same verdict once observational inflation of weak systems is handled). Best-track conventions differ across agencies and eras; the ledger makes this census auditable, the analysis starts at 1980 (full geostationary coverage), and no conclusion rides on the count's absolute level. The pair statistics are isotropic and same-basin; r_eff is an integrated-deficit scale, not a strict minimum separation. The occupancy result is an association with storm-attached controls, not an intervention. The capacity arithmetic in Section 11 is a scaling closure, not a simulation. The utilization ceiling is a counterfactual normalization (all favorable ocean at PI simultaneously), meaningful only as an upper bound — which is the point.

Every number on this page is computed by pipeline/frequency.py from data/archive/, data/structure/, ERSSTv5 and the MPI store, and ships as data/frequency.json + data/frequency_census.csv beside this page.

Section 13

References

1. Sobel, A. H., A. A. Wing, S. J. Camargo, C. M. Patricola, G. A. Vecchi, C.-Y. Lee and M. K. Tippett, 2021: Tropical cyclone frequency. Earth's Future, 9, e2021EF002275.
2. Emanuel, K., 2022: Tropical cyclone seeds, transition probabilities, and genesis. J. Climate, 35, 3557–3566.
3. Hsieh, T.-L., G. A. Vecchi et al., 2022: Model spread in tropical cyclone frequency and seed propensity. Geophys. Res. Lett., 49, e2021GL097157; and 2024: Sci. Adv., 10, eadi2779 (seeds and cloud feedback); Hsieh et al., 2020: Clim. Dyn., 55, 3177–3196; Zhao, H., et al., 2024: Sci. Adv., 10, eadl2142 (the CMIP6 historical decline).
4. Khairoutdinov, M., and K. Emanuel, 2013: Rotating radiative-convective equilibrium simulated by a cloud-resolving model. J. Adv. Model. Earth Syst., 5, 816–825.
5. Chavas, D. R., and K. A. Reed, 2019: Dynamical aquaplanet experiments with uniform thermal forcing. J. Atmos. Sci., 76, 2257–2274. (With Reed & Chavas 2015, the packing/size–frequency prediction tested in Section 6.)
6. Merlis, T. M., M. Zhao and I. M. Held, 2013: The sensitivity of hurricane frequency to ITCZ changes. Geophys. Res. Lett., 40, 4109–4114; Burnett, A. C. et al., 2021: GRL, 48, e2020GL091980.
7. Zhou, Y., and Y. Lin, 2024: A toy model for the global annual number of tropical cyclones. Earth's Future, 12, e2024EF004839; Vu, T.-A., C. Kieu, D. Chavas and Q. Wang, 2021: J. Adv. Model. Earth Syst., 13, e2020MS002207. (The recovery-time models tested in Section 7.) Yang, Y., et al., 2026: Drying effect of landfalling tropical cyclones. Geophys. Res. Lett., 53, e2025GL121284.
8. Frank, W. M., and G. S. Young, 2007: The interannual variability of tropical cyclones. Mon. Wea. Rev., 135, 3587–3598; Elsner, J. B., and T. H. Jagger, 2013: Hurricane Climatology, Oxford University Press. (The prior tests of Sections 2–3.)
9. Schenkel, B. A., 2016: A climatology of multiple tropical cyclone events. J. Climate, 29, 4861–4883; and 2017: J. Climate, 30, 5805–5813. Li, T., and B. Fu, 2006: J. Atmos. Sci., 63, 1377–1389; Krouse, K. D., A. H. Sobel and L. M. Polvani, 2008: J. Atmos. Sci., 65, 644–654. (Multiple-storm separations and Rossby-wave-dispersion genesis, Section 6.)
10. Hoogewind, K. A., D. R. Chavas, B. A. Schenkel and M. E. O'Neill, 2020: Exploring controls on tropical cyclone count through the geography of environmental favorability. J. Climate, 33, 1725–1745. (The packing bound, Section 9.)
11. Hart, R. E., R. N. Maue and M. C. Watson, 2007: Estimating local memory of tropical cyclones through MPI anomaly evolution. Mon. Wea. Rev., 135, 3990–4005; Karnauskas, K. B., L. Zhang and K. A. Emanuel, 2021: The feedback of cold wakes on tropical cyclones. Geophys. Res. Lett., 48, e2020GL091676. (Environmental memory, Section 7.)
12. Ikehata, K., and M. Satoh, 2021: Climatology of tropical cyclone seed frequency and survival rate. Geophys. Res. Lett., 48, e2021GL093626; Levin, E. L., G. A. Vecchi and G. Villarini, 2026: Large climate model ensembles reveal underdispersion in seasonal Atlantic tropical cyclone counts. Mon. Wea. Rev., early online. (Seeds vs survival; the binomial two-stage picture.)
13. Dunkerton, T. J., M. T. Montgomery and Z. Wang, 2009: Tropical cyclogenesis in a tropical wave critical layer: easterly waves. Atmos. Chem. Phys., 9, 5587–5646; Montgomery, M. T., et al., 2012: The PREDICT experiment. Bull. Amer. Meteor. Soc., 93, 153–172. (The marsupial pouch, Section 11.)
14. Davis, C. A., and L. F. Bosart, 2003: Baroclinically induced tropical cyclogenesis. Mon. Wea. Rev., 131, 2730–2747; Moore, R. W., and M. T. Montgomery, 2005: The diabatic Rossby vortex. J. Atmos. Sci., 62, 2703–2725; McTaggart-Cowan, R., et al., 2013: A global climatology of baroclinically influenced tropical cyclogenesis. Mon. Wea. Rev., 141, 1963–1989. (The baroclinic seed pathway.)
15. Chen, L., J. Gao and T. Li, 2024: Formation of long-lasting inactive and active multiple TC events in the western North Pacific. Clim. Dyn., 62, 3823–3840; Fu, Z.-H., et al., 2025: Shifting hotspot of tropical cyclone clusters in a warming climate. Nat. Clim. Change, 15, 850–858; Ito, K., and K. Yamauchi, 2026: Limited influence of pre-existing TCs on subsequent cyclogenesis in the western North Pacific. J. Geophys. Res. Atmos., 131, e2025JD044793. (Confounds and interventions, Sections 6 and 8.)
16. Chand, S. S., et al., 2022: Declining tropical cyclone frequency under global warming. Nat. Clim. Change, 12, 655–661; and the reanalysis-artifact critique, Nat. Clim. Change, 13, 2023.
17. Klotzbach, P. J., et al., 2022: Trends in global tropical cyclone activity: 1990–2021. Geophys. Res. Lett., 49, e2021GL095774.
18. Maue, R. N., 2011: Recent historically low global tropical cyclone activity. Geophys. Res. Lett., 38, L14803.
19. Emanuel, K., 2005: Increasing destructiveness of tropical cyclones over the past 30 years. Nature, 436, 686–688. (The PDI this site's power climatology extends.)