Assessing Future Tropical Cyclone Risk Using Downscaled CMIP6 Projections

Abstract:

The authors employ the Columbia Hazard model (CHAZ) to characterise future tropical cyclone (TC) activity under the shared socioeconomic pathways (SSP) SSP2-4.5, SSP3-7.0, and SSP5-8.5 by downscaling 12 models that participated in the Coupled Climate Model Intercomparison Project’s sixth generation (CMIP6), focusing on the Western North Pacific (WNP) and North Atlantic (ATL) basins. Results from CHAZ are also used in conjunction with the Impact Forecasting Atlantic/Caribbean Tropical Cyclone Wind Model, an industry catastrophe model, to project future changes in financial losses in the ATL. As with previous downscaling of CMIP5 models in CHAZ, projections of TC frequency depend on the choice of moisture variable used in the tropical cyclone genesis index (TCGI), despite similar trends when the two are applied in the historical period. Simulations using column relative humidity (CRH) project an increasing TC frequency trend in the future, while those using saturation deficit (SD) project a decrease. In the WNP, TC frequency scales linearly with the rise in global mean surface temperature, highlighting a direct link with anthropogenic greenhouse gas (GHG) radiative forcing. While ATL TC frequency in the SD experiments exhibits the same trend, the CRH response is complex and nonlinear, probably due to the higher sensitivity of the response of TC potential intensity to aerosol versus GHG forcing. Our projections of financial losses are equally uncertain, consistent with the corresponding bifurcation in TC frequency between the CRH and SD experiments. These projections, despite their inherent uncertainties, can still be useful if viewed as placing bounds on future changes in risk, since not very large increases (i.e., much greater than around 10%) are projected in any of the ATL loss results.

1. Introduction

Tropical Cyclones (TCs) are among the most devastating natural hazards. A large and growing number of studies estimate the statistics and impacts of potential future TCs, supported by a wide range of modelling and empirical techniques. One of these techniques is catastrophe modelling. TC catastrophe models adopt a probabilistic approach, simulating large numbers of synthetic events to assess the likelihoods and severities of potential storms and their associated losses. Catastrophe model output can be presented in the form of risk maps, metrics of spatially aggregated losses, and other information that can be readily used by insurers, reinsurers, emergency management agencies, and other stakeholders in pricing insurance policies, allocating resources for disaster response, determining capital reserves, and developing risk management strategies (Grossi and Kunreuther 2005, Vickery et al. 2000.

TC catastrophe models consist of multiple components. The hazard assessment component consists of the synthetic TCs themselves, including wind fields, and in some cases precipitation and storm surge flooding. The exposure and vulnerability components describe the assets at risk and the extent to which a given hazard (wind, flood depth, etc.) intensity will damage them. These inform estimates of financial losses. Additionally, uncertainties in the hazard component propagate downstream, affecting the final loss estimates, thus emphasising the importance of accurately modelling the initial hazard. Most traditional catastrophe models have built their hazard catalogues directly on historical data, using statistical methods to generate synthetic events close to those data, without explicitly considering the potential influence of trends in the historical record. To assess the influence of human-induced climate change, in particular, some degree of physical modelling of how climate change affects TCs is needed. Here, we present a study in which a statistical dynamical model is used to downscale physically-based future climate projections to generate synthetic North Atlantic and western North Pacific TCs in a form amenable to catastrophe modelling. The results are then used in conjunction with an industry catastrophe model to project future changes in losses.

Statistical-dynamical TC downscaling models (e.g., Emanuel et al. 2008, Lee et al. 2018, Bloemendaal et al. 2020, Jing and Lin 2020, Lin et al. 2023) use reanalysis or climate model data as inputs representing environmental variables relevant to the genesis, tracks, and intensities of TCs. By taking the necessary environmental data from climate models, the effects of climate change can be represented. Unlike climate models, such downscaling models are computationally efficient and can simulate large numbers of storms at a much lower cost than the high-resolution global dynamical models that are used in many studies of TCs and climate. Being partly empirical, it can also be easier (compared to a fully dynamical model) to ensure that a statistical-dynamical model has TC frequencies and intensities at landfall that to some extent agree with historical observations, as is important in practical applications. Such models have become commonly used tools for quantifying TC risk and a vital part of the input datasets needed to estimate TC-induced losses in catastrophe models (Meiler et al. 2022, Aznar-Siguan and Bresch 2019, Baldwin et al. 2023).

Lee et al. (2020) used the Columbia HAZard model (CHAZ, Lee et al. 2018), a statistical dynamical downscaling model, to downscale future TC activity in simulations of six models from the phase 5 of the Coupled Model Intercomparison Project (CMIP5; Taylor et al. 2012).Among other results, they found that TC activity would either increase or decrease with global warming, depending on whether CRH or SD was chosen as the humidity variable in the genesis component of the model, respectively. This study builds on the work of Lee et al. (2020), applying CHAZ to generate TC projections from 12 models in phase 6 of the Coupled Intercomparison Project (CMIP6; Eyring et al. 2016). Besides using the newer generation of climate models in CMIP6 as opposed to CMIP5, we extended and modified Lee et al. (2020) in several ways, including doubling the size of the CMIP multi-model ensemble being downscaled and implementing an improved bias correction scheme. Perhaps most importantly, the results are used to modulate historical financial loss estimates using the Impact Forecasting Atlantic/Caribbean Tropical Cyclone Wind Model, released in 2019 to produce estimates of future changes in TC losses.

2. Methods and Experiment Design

2.1 The Columbia HAZard model

The Columbia HAZard Model (CHAZ, Lee et al. 2018) is a statistical-dynamical downscaling model designed to establish a physical link between large-scale climate drivers and tropical cyclones. CHAZ consists of three distinct components that predict genesis, storm track, and intensity. The genesis model generates weak disturbances with a rate determined by environmental conditions, calculated using the Tropical Cyclone Genesis Index (TCGI) of Tippett et al. (2011) as later modified by Camargo et al. (2014). The TCGI calculation requires various environmental variables, namely tropical cyclone potential intensity (PI, Bister and Emanuel 2002), vertical wind shear, a humidity variable, and the 850-hPa absolute vorticity. These variables determine the seeding rate, 𝜇_MV, as follows:

The moisture variable MV is either CRH or SD. As in Lee et al. 2020, we refer to the two resulting estimates as TCGI CRH and TCGI SD. The coefficients 𝑏 are constants, 𝜂 is absolute vorticity at 850 hPa, PI is potential intensity (Bister and Emanuel 2002), and SHR is vertical wind shear, calculated as the magnitude of the vector wind difference between 200 and 850 hPa.

The coefficients for the TCGI discussed in this study have been updated from previous works by Lee et al. (2018) and Lee et al. (2020). Unlike the earlier versions, which were derived from a Poisson regression using environmental fields from the European Centre for Medium-Range Weather Forecasts (ECMWF) ERA-Interim reanalysis (Dee et al. 2011), the current coefficients are based on the latest ECMWF reanalysis version - ERA5 (Hersbach et al. 2020), calculated over the period 1981-2010. For TCGI CRH, the coefficients are b= -24.132, b_η=2.512, b_CRH=0.077, b_PI=0.062, and b_SHR= -0.120; and for TCGI SD, b= -18.356, b_η=2.483, b_CRH=0.073, b_PI=0.080, and b_SHR= -0.135. Some climatological features of the TCGI based on Poisson regression of the ERA5 reanalysis environmental fields, as compared with other reanalysis products for the common period 1980 – 2016, are shown in Dirkes et al. (2023).

The track component of CHAZ is based on a beta and advection model that moves storms forward via advection by the environmental wind plus a beta drift component (Emanuel et al. 2006). The model also includes a stochastic component based on eddy statistics to account for the effects of submonthly wind fluctuations. The intensity component is based on a multiple linear regression on key environmental parameters, in addition to an autoregressive stochastic error term. The latter accounts for the internal storm dynamics and does not depend explicitly on the environment, but rather on the storm's current state and recent history. A separate regression model is employed to estimate the intensity of the storm during landfall and in the immediate aftermath, factoring in both proximity to land and environmental conditions (Lee et al. 2015, 2016, 2018).

2.2 Downscaling TC Activity In CMIP6 Models

We downscaled CHAZ from 12 models that participated in CMIP6 (Table 1). We used simulations from the historical scenario (HIST, 1951-2014) and three future scenarios based on three shared socio-economic pathways (SSP), namely SSP2-4.5, SSP3-7.0, and SSP5-8.5, covering the years 2015-2100. While the downscaling was global in scope, we focused our analysis on TC activity in two basins – the Western North Pacific (WNP, 0-45N,100-180E) and the North Atlantic (ATL, 0-45N, 20-100W). By juxtaposing the historical period with projections from each of the three SSP scenarios, and considering two basins and two humidity variables, we generated 12 distinct output datasets for each of the 12 models.

For each CMIP6 model and scenario (HIST, SSP2-4.5, SSP3-7.0, SSP5-8.5), we downscaled a single CMIP6 ensemble member in CHAZ. In each instance, we applied 10 realisations of random seeding (resulting in different genesis locations and thus tracks) and 40 realisations of the stochastic intensity model, making it possible to simulate 400 combinations of synthetic TC event sets. Consequently, for each of the 12 output datasets, we produced a total of 25,600 years' worth of synthetic TC data in the historical era, corresponding to 64 years (1951-2014) multiplied by 400 combinations and 34,400 in the future, that is 86 years (2015-2100) multiplied by 400 combinations. Each set of simulations constitutes a dataset that includes at least 10,000 storms. Such a scale is often necessary for assessing TC risk in any specific region, as suggested by Lin and Emanuel (2016). The resulting synthetic tracks contain information on the storm locations in longitude and latitude at six-hour intervals, as well as the maximum wind speed measured in knots (kt).

In a new step since Lee et al. (2020), we applied bias correction to the climatology of the environmental fields produced by the CMIP6 models before passing them to CHAZ’s genesis component. The differences between the monthly historical model climatology and the ERA5 reanalysis climatology were estimated for each variable over the period of 1981–2010. These monthly differences were used to adjust the historical model climatology toward the ERA5 climatological fields. This step greatly reduces the biases in the TC statistics produced by CHAZ as compared to historical observations. The same monthly climatological differences were added to the climatological fields in future scenarios, based on the assumption that model biases would remain constant. We did not modify the steering flow and environmental predictors for the CHAZ's track and intensity components. However, in CHAZ, environmental predictors for intensity calculation are normalised by each model's tropical mean, which, in itself, is a way to remove mean biases in the CMIP6 models.

2.3 TC Basins and Analysis

Within our target basins, we focus on the statistics of storms crossing specific landfall “gates.” We defined 20 gates in the WNP and seven in the ATL for US Atlantic coastal regions, respectively (Fig. 1). In order to delineate these gates, we used the Natural Earth coastline physical vector dataset, which offers high-resolution data at a 1:50 million scale available at naturalearthdata.com (Kelso and Patterson 2010). This translates to a roughly 25 km grid resolution along the coastline allowing for a more accurate representation of TC landfalls at the local scale. We identified landfalling TCs and their corresponding locations by determining the intersection points between the synthetic TC tracks in the CHAZ output and the coastline segments within each gate.

Figure 1: The geographical context for the study, highlighting selected tropical cyclone risk regions (i.e., TC gates): 7 in the Atlantic Basin and 20 in the WNP Basin. Superimposed in kts are [a] Tropical cyclone tracks from the International Best Track Archive for Climate Stewardship (IBTrACS USA, since 1951) and [b] synthetic tropical cyclone tracks from CHAZ based on the historical run of CESM2 (1951-2014, mean of CHAZ’s 40-member intensity ensemble).

In all our analyses, we compute TC statistics on an annual basis and our analysis of the historical era focused on the time period from 1980 to 2014; 1980 corresponds approximately to the start of the satellite era, while the historical simulations end in 2014. For the future projections, we considered three epochs: the near future or early-century (2030–2050), mid-century (2060–2080), and late-century (2080–2100).

To evaluate how well the selected climate models capture key TC characteristics of the observed record, we compared them to the International Best Track Archive for Climate Stewardship (IBTrACS, Knapp et al. 2010) dataset version 4 – a compilation of information from multiple international meteorological agencies, which provides updated information on TC tracks, intensities, and other attributes every three hours. We used the version of the data in which the best tracks are contributed by the U.S. agencies, the National Hurricane Center (NHC) for the Atlantic and the Joint Typhoon Warning Center (JTWC), for the WNP, hence the designation IBTrACS USA.

2.4 Estimation of Landfalling TC Risk

To quantify the TC hazard at each gate, we employed both annual frequencies of exceedance and their corresponding return period curves for landfalling TCs across various intensity categories for both the observed and modelled storm tracks. First, we calculated the probability density function (PDF) of TCs based on the Saffir-Simpson wind scale and determined the annual frequency of landfalling TCs with wind speeds reaching values equal to or greater than 40 kt at each gate. This allowed us to create a frequency distribution at different wind speed values for each gate. The return period for a specific intensity level was then calculated as the inverse of its corresponding annual exceedance probability, i.e., the likelihood of one or more TCs making landfall with an intensity greater than that level in a given year.

From the TC landfalling frequency distributions, we computed the changes in the frequencies of landfalling TCs in each intensity category resulting from anthropogenic climate change at each gate. We refer to these as climate change deltas (cc-deltas), calculated by subtracting the mean landfalling TC frequency (i.e., average of all seeding rates and intensity ensemble for the 12 models), defined by epoch, of the future climate scenarios from the mean historical frequency and normalising by the latter to produce a percentage change. We assess the impact of anthropogenic climate change on other TC properties by following the cc-delta approach, thus comparing the climatologies of a relevant parameter in future and historical periods and normalising the outcome by the historical value. To assess the statistical significance of the changes, we applied a two-sided Student’s t-test to the combined 12-member model, each comprising a 10- and 40-member seeding and intensity ensemble, with a p-value of 0.05 serving as the threshold for significance.

2.5 Estimation of Losses

Financial losses were calculated using the Impact Forecasting Atlantic/Caribbean Tropical Cyclone Wind Model (hereafter IF model), focusing only on the ATL. The IF model is a catastrophe model designed to provide probabilistic estimates of losses from hurricanes on given insured assets. It consists of four modules: hazard, vulnerability, exposure, and financial. Input information into the catastrophe model is called a portfolio, here a generic one that reflects industry exposure, and contains a list of insured properties with their locations (latitude/longitude, ZIP code, County, etc.), insured sums with insurance policy conditions, and some description of the structure that the model can understand (occupancy type, construction type, height of the structure, year it was built, etc.) and that is relevant to the degree of damage caused by high winds. The output of the catastrophe model is a table with estimated damages expressed monetarily for a set of synthetic events representing the hurricane activity in the United States and can be used to plot exceedance curves depicting expected monetary loss incurred on a given portfolio as a function of return period (inverse of cumulative probability) in years.

The hazard module of the IF model represents 50,000 years of TC activity in the North Atlantic Basin and consists of synthetic events modelled from event genesis through dissipation using statistical regression models trained upon the HURDAT2 dataset (1850 through 2017). First, the number of events per each stochastic year is randomly generated from a Poisson distribution, and then the initial storm positions are randomly drawn from a spatial probability map which reflects the general potential index (Camargo, 2007) – a proxy for the spatial distribution of the first occurrence of tropical cyclones across the basin. Once the storms are seeded, their movement and intensity changes are guided by regression models, which are trained separately for each 5 by 5-degree region covering the entire basin (Vickery et al., 2000; Darling, 1991) using observed HURDAT2 track parameters within that 5 by 5-degree region. When a simulated track makes landfall, the intensity model is replaced by a filling decay model (Vickery, 2005) to reflect rapid intensity changes after the storm loses its source of energy.

To complete the set of parameters needed to create wind fields, the radius of maximum wind (Rmax) is estimated. The Vickery and Wadhera (2008) implementation is used, relating Rmax to the intensity and latitude of the storm. For each storm, Vmax, Rmax, and forward motion are used to generate a spatial wind field following Willoughby et al. (2006), which is characteristic of idealised, open terrain conditions. The idealised wind model is then adjusted to estimate three-second gusts and to account for upstream changes in terrain roughness (ESDU, 2002a). This way, the model has an estimate of gust speeds for land locations with 1 km resolution for a large number of hurricanes spanning thousands of years.

The vulnerability module of the catastrophe model is a set of vulnerability curves that link damage ratios – the monetary estimate of damage on a structure relative to the replacement value of the entire structure – to the hazard, which in this case is the three-second gust. These functions are estimated using a performance-based hurricane engineering framework with multi-layer Monte Carlo simulation for the loss analysis of residential buildings subject to hurricane hazard (Unnithan et al., 2016). Calibration using insurance claims is then applied and is necessary to align the engineering estimate with the damage experience. The exposure module in the catastrophe model serves as a dictionary translating the information about location (coordinates, ZIP codes, etc.), detailed information on structure occupancy, construction type, building height, year built, and insurance policy conditions (insured value, limits, deductibles, etc.) into information the model can correctly interpret and use for the loss estimate. For the provided location and structure information, the module assigns identification numbers that the model uses to find the corresponding hazard and vulnerability values in large files with data.

The model loss calculation is conducted as a Monte Carlo process where random sampling from known distributions (to account for uncertainties in structure location and damage ratios) leads to multiple loss estimates per structure. Lastly, the financial module applies provided policy conditions to obtain the estimated liability of the portfolio components given policy conditions such as deductibles and limits (that is, the estimated payout from the insurance company to the insured person exceeding the deductible but capped by the policy limit). The losses are aggregated to the required level, which is typically the portfolio level, resulting in an estimate of the mean loss an insurance company would incur over a large number of hypotheticals, i.e. simulated hurricanes that realistically represent hurricane activity in the basin.

Our interest here is in the projected changes in those losses due to climate change. To achieve this, we used a frequency adjustment approach in this study: for each intensity category and each gate, the cc-deltas from CHAZ were applied to the frequencies of the corresponding IF model synthetic storms. In the case of IF model tracks with multiple landfalls, the frequency was adjusted based on the cc-delta of the related category and gate. This was done for only the SSP2-4.5 and SSP3-7.0 scenarios. In comparison to SSP5-8.5, the SSP2-4.5 and SSP3-7.0 scenarios are more moderate emission scenarios that better align with various global policy targets (Rogelj et al. 2018). Percentage changes in multiple statistics of financial loss were derived from the modelled stochastic (current) losses in the IF model modulated by the CHAZ-derived cc-deltas.

3. RESULTS

3.1 Changes in Annual TC Frequency

We first compare the frequencies of simulated TCs during the historical period to those in observations (Fig. 2). For both the WNP and the ATL, the mean TC frequencies in the ensemble approximately match those seen in observational data. This is probably a consequence of the bias corrections applied to the model environmental fields, as well as the fact that CHAZ was trained on historically observed storms (Lee et al. 2018). In the Atlantic, the TC frequency trends in the observed historical period appear larger than those of the simulated ensemble means, but the observations remain within the interquartile range of the model results, except for the couple of years with the greatest frequencies (Lavender et al. 2018). Attribution of recent trends in TC frequency to specific causes remains under debate, however. In the Atlantic for instance, any possible greenhouse gas-driven trends can be obscured by both natural climate variability and anthropogenic aerosol forcing, as well as changes in observation methods (Emanuel 2021, Vecchi et al. 2021, Chand et al. 2022, Klotzbach et al. 2022).

Figure 2: Time series of CHAZ-simulated annual-mean TC frequency. Thin lines in the orange (blue) shade of colours represents individual outcomes from the participating CMIP6 models (Table 1), while the thick red and blue lines show the ensemble mean. The shading around the mean represents the 25th to the 75th percentile range. The observed TC frequency is superimposed in the thick black line.

In future projections, the changes in projected TC frequency are highly sensitive to the choice of humidity variable in CHAZ’s genesis component, echoing the findings of Lee et al. (2020). Specifically, when CRH is used as the moisture variable in CHAZ's genesis model, we observe projections of increased TC activity. Conversely, using SD leads to projections of decreased TC activity. This divergence occurs, despite both metrics, each designed to measure atmospheric moisture, producing similar trends in the historical climate (Camargo et al. 2014, Lee et al. 2020; 2022; 2023).

In the WNP, the magnitudes of TC frequency trends, increasing for CRH and decreasing for SD, scale with global mean surface temperature changes under different SSP scenarios (Fig. 2; Fig. S1). The ATL shows some similarities, particularly a comparable decline in TC frequency in the TCGI SD experiments (Fig. 2). However, unlike the WNP, the ATL's TCGI CRH experiments show almost no trends in any scenario, suggesting that the magnitude of the trend does not correlate with global mean surface temperature. To help elucidate these TC frequency changes, the time series of TCGI for each experiment is shown in Fig. 3. TCGI directly controls TC frequency in CHAZ, resulting in similar behaviour between the two. Again, the TC frequency trends in the Atlantic under CRH are very small, showing almost no increase in the SSP2-4.5 and SSP3-7.0 scenarios.

Figure 3: Time series of mean TCGI for one ensemble member each from the 12 participating CMIP6 models (same members downscaled in CHAZ). Thin dotted lines in the orange (blue) represent individual ensemble members for CRH (SD). The ensemble mean is shown in thick red for CRH and thick blue for SD. The observed TCGI is superimposed in the thick black line.

3.2 Track Density and Spatial Distribution

To illustrate the spatial patterns in the evolution of TC activity, we compared the ensemble mean TC track density differences between the future and historical periods (Fig. 4). Track density is determined by counting the number of times per year that the TC tracks pass through a 2° x 2° grid cell. Counting is done at six-hour intervals, consistent with CHAZ’s track output. For reference, the historical track density is superimposed on each plot in contours.

Figure 4: Mean track density differences between the future (2015-2100) and historical (1951-2014) periods. Historical track density (first panel) is superimposed on the difference plots in contour. Grey dots show that the track density differences are statistically significant at p = 0.05 based on a two-sided Student’s t-test from the 12-model, 10 random seeding realisations and 40-member intensity ensemble.

Changes in Fig. 4 are consistent with the frequency plots in Fig. 2. Across the WNP, track density increases in the TCGI CRH experiment and declines in the TCGI SD experiment, and the higher the radiative forcing associated with the scenario, the stronger the degree of change. In the ATL, again, changes between SSP2-4.5 and SSP3-7.0 are not discernible. Globally, in CRH, we see a northward shift of TC activity, with increasing activity in the WNP and northernmost ATL basins. In the Southern Hemisphere, there is decreasing activity in the South Indian and South Pacific basins but increasing activity in the Australian basin.

Given that the WNP generates a substantially higher number of storms per year on average compared to the ATL (approximately 25 in the WNP versus 12 in the ATL, Fig. 2), we also calculate normalised track density changes to enable an objective comparison of the two basins, and also allow us to detect regional shifts in TC occurrence and distribution while removing the global mean trend (Fig. 5). Normalisation is achieved by dividing the track density at each grid point by the global sum of the annual mean track density over all grid points. This is done individually for each model and epoch (i.e., 2030-2050, 2060-2080, and 2080-2100), effectively ensuring that the global mean trend is removed. The normalised data are then used to compute the multi-model mean track density.

Figure 5: Normalised mean track density differences between the future (2015-2100) and historical (1951-2014) periods. Normalisation is achieved by dividing each epoch by the sum of its global TC number before taking the difference. Historical track density (first panel) is superimposed on the difference plots in contour. Grey dots show that the track density differences are statistically significant at p = 0.05 based on a two-sided Student’s t-test from the 12-model, 10 seeding rates and 40-member intensity ensemble.

The resulting patterns are remarkably independent of the epoch, as well as of the choice of humidity variable, demonstrating that the global mean trend and spatial pattern are controlled by different aspects of the inputs. The ATL shows a dipole-like pattern in track density changes, i.e., a decline in the Gulf of Mexico and an increase in the Atlantic, which happens in both moisture experiments. There is also a slight decrease in storm activity over the South China Sea in the WNP, as well as the southern hemisphere decreases noted above. At least in the northern hemisphere, the geographic shifts shown in Fig. 4 and Fig. 5 appear broadly consistent with the well-documented poleward migration of TCs with warming, especially in the WNP (Kossin et al. 2014, 2016, Nakamura et al. 2017, Daloz and Camargo 2018, Lin et al. 2024).

3.3 Basinwide TC Frequency

In Table 2 we use the climate change delta approach, as outlined in the methods section, to quantify basinwide TC frequency changes for all landfalling TCs. In the WNP, TC frequency generally increases with epoch and under increased radiative forcing in the CRH experiment. Conversely, the SD experiment reveals a basinwide decline in TC frequency in response to increased radiative forcing, consistent with the previously observed divergence in TC frequency. A similar pattern is seen in the ATL SD experiment, but not in its CRH counterpart, where changes are not consistently correlated with time or radiative forcing.

Because TC wind hazard is influenced more by the occurrence of the strongest storms than by the overall frequency of TCs, we quantified basinwide TC frequency changes in each Saffir-Simpson intensity category (Fig. 6). The results follow the cc-deltas presented in Table 2. The WNP TCGI CRH experiment exhibits a consistent pattern, with TC frequency increasing across epochs and at all intensity levels. In the WNP TCGI SD experiments, TC frequency decreases over time. In both instances, the magnitude of change becomes larger with increasing radiative forcing. These variations with intensity show that, in general, global warming increases TC intensity across both the CRH and SD results, although in the CRH results, frequency increases more for more intense storms, while in the SD results, frequency decreases less for the more intense storms.

Figure 6: Basinwide climate change deltas. The black vertical lines are the 95% confidence interval from the 12 model, 10 seeding rate, and 40-member intensity ensemble.

The ATL results are again more subtle to interpret. It remains true that in the CRH result, the greatest increases in frequency occur at the highest intensities, but the changes do not generally increase with time or with radiative forcing, particularly for the SSP2-4.5 and SSP3-7.0 experiments. In the SSP2-4.5 scenario, TC frequency actually declines across epochs, particularly for the lower-category storms. Consistent with results above, TC frequency in the ATL SD experiments declines consistently with time as radiative forcing increases, and in terms of magnitude, represents the most dramatic changes in comparison to the CRH result. Interestingly, for categories 4 and 5, the SD frequencies increase in the early future in SSP2-4.5 before decreasing later.

Projected changes in TC frequency by categories are also computed at the seven ATL gates and seven of the most active WNP gates (Fig. 7). In the WNP, the cc-deltas at individual gates align largely with the basinwide TC changes shown in Fig. 6, especially in the TCGI CRH experiment. The TCGI SD experiment has a few exceptions, namely the Mid-China, SouthJPN (Southern Japan), and EastJPN (Eastern Japan) gates, where TC activity is projected to increase at several intensity levels, most notably in SSP2-4.5. These regional variations in TC frequency changes align with our previous result regarding the decrease in storm density in and around the vicinity of the South China Sea, as illustrated in Fig. 5.

Figure 7: Climate change deltas by gate, i.e., the difference between referenced future periods and the historical period. Changes are calculated over three future epochs, i.e., 2030-2050, 2060-2080, and 2080-2100, represented by a triad of columns on the x-axis and from left to right respectively. TC intensity categories are shown on the y-axis. TS is a tropical storm.

The most notable disagreements between basinwide frequency changes and those at individual gates occur in the Atlantic Basin. In the CRH experiment, increasing TC frequency in the Southeast, Mid-Atlantic, and Northeast gates agree with the expected basinwide change, while eFlorida (East Florida) and the gates along the Gulf Coast see some declines in TC frequency. However, even when the cc-deltas share the same sign in both the SD and CRH results, the projected changes still diverge significantly, maintaining the bifurcation in their projections regardless of the sign of change. In the ATL TCGI SD experiment, the opposite pattern emerges, where increases in the frequency of the most intense storms do occur at the Southeast, Mid-Atlantic, and Northeast gates, primarily but not exclusively in SSP2-4.5. Here as well, our results are congruent with the projected decrease in storm density in the Gulf of Mexico and the simultaneous increase in the ATL (Fig. 5).

3.4 Landfall Return Period on Select Gates

We used the synthetic TC tracks to derive return periods at the selected seven WNP gates (Fig. 8 a,b) and all seven ATL gates (Fig. 9 a,b). Given the considerable number of gates, scenarios, and time periods available, we focus specifically on the late-century period in SSP5-8.5, juxtaposed with the historical and observed return period curves.

Figures 8a and 8b: IBTrACS and return period for selected gates in the Western North Pacific, including best-track (black lines) over the period of 1951-2022, and from CHAZ synthetic storms tracks over the historical period (1951-2014) and late-century in SSP5-8.5 (2080-2100). The grey shadings show the spreads of the return period curves from individual ensemble members from historical period while the solid grey lines are the return period curves using all data. The blue and coral shadings and the respective solid lines are return period curves and the model spread from future climate scenario. Numbers at each Saffir-Simpson intensity threshold are the percentage changes of the frequency of the storms exceeding the threshold.

Comparing the observed and simulated historical return period curves reveals no definitive pattern from which we can ascertain whether TCGI CRH or TCGI SD is more closely aligned with the historical observed curve. Overall, the historical return period curves for both TCGI CRH and TCGI SD display closer agreement with the observation-based ones in the WNP (Fig. 8 a,b) than in the ATL (Fig. 9 a,b). However, the results for West Florida (wFlorida), Northeast, and Southeast US exhibit particularly close alignment with observations. CHAZ tends to underestimate the historical frequency of high-intensity storms at the other ATL gates. To improve the realism of the return period curves at the regional scale, it may be beneficial to explore localised calibration approaches, while being cautious of over-fitting (Lee et al. 2022).

Figures 9a and 9b: IBTrACS and return period for all Atlantic gates, including best-track (black lines) over the period of 1951-2022, and from CHAZ synthetic storms tracks over the historical period (1951-2014) and late-century in SSP5-8.5 (2080-2100). The grey shadings show the spreads of the return period curves from individual ensemble members from historical period while the solid grey lines are the return period curves using all data. The blue and coral shadings and the respective solid lines are return period curves and the model spread from future climate scenario. Numbers at each Saffir-Simpson intensity threshold are the percentage changes of the frequency of the storms exceeding the threshold.

In the future SSP5-8.5 scenario, projections from the CRH experiment in the WNP (Fig. 8 a,b) indicate that most gates will experience decreases in landfalling TC return periods at fixed intensity thresholds, signifying increases in hurricane hazards at these thresholds. In contrast, the SD experiment anticipates increases in these return periods, suggesting a future decrease in hurricane hazard across all intensity thresholds.

Unlike the WNP, not all the ATL gates (Fig. 9 a,b) show a decrease in return periods in the CRH result. This is particularly true for the gates in the Gulf of Mexico; for example, the return period at the Texas gate increases, especially for lower-intensity storms. Overall, the magnitude of the changes in the return periods in the ATL SD experiment is significantly larger than those observed in its CRH counterpart or even the WNP. These findings are consistent with the illustrated changes in track density (Fig. 5), as well as the basinwide (Fig. 6) and gate-specific (Fig. 7) TC frequency changes.

3.5 Assessment of TC Frequency and Change Probability in CMIP5 versus CMIP6

The study by Lee et al. 2020, using CMIP5 data to downscale with CHAZ, provides a reference to which we can compare the TC projections based on CMIP6. As shown in Fig. 10, we contrast the annual mean TC frequency and relative probability in CMIP5 and CMIP6 for the strongest forcing scenarios, RCP8.5 and SSP5-8.5 scenarios. The CMIP5 plots are adapted from Lee et al. 2020, who used an ensemble of six models in their research, namely CCSM4, GFDL CM3, HadGEM2-ES, MPI-ESM- MR, MRI-CGCM3, and MIROC5. In contrast, our CMIP6 ensemble comprises 12 models. In addition, the simulations in CMIP5 and CMIP6 also differ in their historical periods, with CMIP5 concluding in 2005 and CMIP6 extending to 2014. Although the two sets of models are not identical, all six models from CMIP5 are represented in some way in our CMIP6 ensemble, either as direct predecessors or as slightly modified versions.

Figure 10: Comparing results from CMIP5 to CMIP6: [a, b] Time series of annual-mean frequency for the Atlantic and WNP Basins. Thin dotted lines represent individual ensemble members while the solid lines represent their mean. [c, d] Relative probability of the annual TC frequency from observations (black solid line), HIST (coloured dashed line), and Future strongest forcing RCP8.5 and SSP5-85 scenarios (coloured solid line). Results are based on synthetic storms generated from six CMIP5 models versus 12 from CMIP6 (Table 1).

As expected, Fig. 10 shows a clear divergence in mean TC frequency between the TCGI CRH and TCGI SD experiments in both CMIP5 and CMIP6. The TCGI CRH frequency trends are comparable in the ATL (Fig. 10a) and WNP (Fig. 10b), although CMIP5 shows a mild increase over the projected period. The TCGI SD trends, on the other hand, show a more dramatic difference between CMIP5 and CMIP6, with the former showing a much steeper decline in TC frequency and more so in the WNP than the ATL. The right panels, Figs. 10c and 10d show the distribution of the annual TC frequency from observations and simulations from the historical period (HIST) and future RCP85 and SSP5-8.5 scenarios. For each curve, the mean is marked by a solid line. The means of the HIST results from both CMIP5 and CMIP6 are adjusted to the observations over the common period of 1981 to 2005; consequently, by construction, they exhibit the same value as the observations. This adjustment only has an aesthetic effect, making it easier to compare and interpret the probability distribution of current versus future TC frequency changes.

The simulated historical distributions (coloured dashed curves) are close to those from observations (black curves) in both basins, indicating that CHAZ successfully reproduces the differing observed interannual frequency. In the WNP, CHAZ overestimates the probabilities in the tails of the distributions of the formation rates in both CMIP5 and CMIP6. The distributions of the ensemble mean annual TC frequency in the RCP85 from the TCGI CRH shift toward the right relative to the historical period (Fig. 10, comparing the dashed blue lines to the solid blue lines in the right panels), consistent with the increasing trends in frequency in those experiments.

In the WNP CRH experiment (Figs. 10b and 10d), the ensemble mean annual TC frequency increases by a rate of 0.12 storms per year in CMIP6, representing a mean annual TC frequency increase from 25.5 to 31.9 and a 25% increase by the end of the century. There is a comparable ensemble mean increase by a rate of 0.13 storms per year in CMIP5, equal to a mean frequency increase from 25.5 to 32.3, thus a 27% increase by the end of the century. Projections among individual models, on the other hand, vary significantly, ranging from a 12% decrease in CMIP6 and 8% in CMIP5 to a 78% increase in both datasets. Conversely, the TCGI SD experiment reveals a decrease in the ensemble mean frequency at rates of 0.09 and 0.17 in CMIP6 and CMIP5, corresponding to a decrease from 25.5 storms per year in the current climate to 20.8 by the end of the century in CMIP6, an 18% decrease, and an even more pronounced dip in CMIP5 to 15.8, a 38% decline. Again, the results from individual models yield a broad spectrum of outcomes. In CMIP6, projections range from 8.5 to 29.5 storms per year, indicative of a 67% reduction to a 15.7% increase. In CMIP5, the range is from 9.6 to 24.5, representing a decrease between 63% and 3.9%.

In comparison to the ATL (Figs. 10a and 10c), the ensemble mean rate of change in TC frequency is the same across datasets in each experiment. The CRH experiments show an increase of 0.01 and the SD experiments exhibit a decrease of 0.08. The CRH result shows no change in the ensemble mean TC frequency from the current climate, i.e., 11.9 storms per year, to about 12 at the end of the century. This is the case in both CMIP6 and CMIP5. Likewise, the inter-model spread is similar in both datasets, spanning from 7.5 storms per year and a decline of approximately 37% to around 20 storms per year, a 68% increase. In the SD experiment, the mean annual TC frequency decreases from 11.9 storms per year to 7.4 in CMIP6 and to 5.5 in CMIP5 by the end of the century. These represent decreases of 38% and 54% in CMIP6 and CMIP5, respectively. Here, the inter-model spread is more pronounced, starting from 4.5 to 12.5 (i.e., a 62% decline to almost no change) in CMIP6 and from 3.5 to 8.5 in CMIP5 (i.e., a 71% to 29% decline).

The lower ensemble mean TC frequency in the CMIP5 SD experiment compared to CMIP6 is noteworthy, as are the variations in the ensemble spread between the two. One potential reason is the difference in ensemble sizes, with CMIP5 utilising six models and CMIP6 utilising 12. Additionally, CMIP6 models exhibit a wider range of climate sensitivity compared to CMIP5 (Meehl et al. 2020), resulting in more varied climate feedback responses. However, the accuracy of projections from models with significantly higher climate sensitivities is still debated. Some studies suggest that high-sensitivity models struggle to accurately reproduce historical temperatures and conditions from the last ice age (Ribes et al. 2021, Tokarska et al. 2020, Brunner et al. 2020). Despite this, the ensemble reflects the state of the art, encompassing models across the full range of climate sensitivities.

Meanwhile, the mean rate of change in TC frequency in the CRH experiments is consistent across the datasets. Relative humidity is the ratio of the mass of water vapour in a given mass of air to its saturation value, while SD is the difference of the same two values. Sobel et al. (2019) explained that in a warming climate, as relative humidity in the tropical lower troposphere remains close to constant in most models (e.g., Sherwood et al. 2010), the saturation deficit increases because it is a fixed fraction of an increasing saturation specific humidity. Thus, SD varies more strongly with temperature, leading to greater variations in projections as different models project different temperature trends.

Despite these disagreements, the epistemic uncertainty in TC frequency in CMIP6, in which TC frequency increases or decreases depending on which humidity variable is used in CHAZ’s environmental index for the probability of cyclogenesis, is consistent with CMIP5. This echoes Lee et al. (2020, 2022, 2023), who detailed and synthesised the challenges in discerning future TC frequency trends in the seven TC basins globally, and specifically in the North Atlantic basin and the Northeast Atlantic. Their analysis utilised, among other data products, synthetic storms downscaled from five of the aforementioned CMIP5 models using CHAZ. In the final analysis, they suggest that such uncertainty might be better understood through a 'storyline approach' as proposed by Shepherd et al. (2018), or by considering each of the two contrasting frequency projections as a plausible scenario.

3.6 Financial Losses

Figure 11 shows the projected changes in several statistics of financial losses for the United States (i.e., ATL). Specifically, it illustrates the percentage changes in expected losses plotted for nine return periods: the annual average loss (AAL), the 10-year, 25-year, 50-year, 100-year, 200-year, 250-year, 500-year, and 1000-year return periods. These return periods are assessed across select epochs, representing losses from the early century (2040), mid-century (2070), and late-century (2090). For clarity and ease of reference, the actual loss percentages are provided in Table 3. In both the SSP2-4.5 and SSP3-7.0 scenarios, the progression of loss estimates mirrors the simulated changes in TC frequency, demonstrating that the sensitivity to the choice of humidity variable in CHAZ's genesis component propagates from TC frequency (Fig. 2) through to loss.

Figure 11: Estimated changes in insurance property losses in the United States as per Impact Forecasting Atlantic/Caribbean Tropical Cyclone Wind Model. The horizontal axis describes AAL – Annual Average Loss and return periods of 10, 25, 50, 100, 200, 250, 500 and 1000 years. Solid lines represent the mean change, while dashed lines represent changes from individual models. Red indicates results using column relative humidity (CRH) while grey represents results using saturation deficit (SD), as described in the text. The shading around the mean represents the 25th to the 75th interquartile range.

The 2040 losses from the SSP2-4.5 CRH experiment reveal an initial 25% increase, which reduces to around 5% at the 100-year return period and diminishes further to about 3% at 1,000 years. The corresponding SD experiment shows a contrasting trend, beginning with a 5% reduction in losses and then to a break-even point at the 100-years through to the 1,000-year return period. In this context, a break-even point (0% loss) denotes a situation where the anticipated loss equals the historical baseline loss. The projected trajectory of losses during the early century (2040), particularly in the CRH experiment, remains largely unchanged throughout the mid (2070) and late (2090) centuries. By contrast, the SD scenario shows a progressive decline in losses at each return period, consistent with a corresponding decline in TC activity with time (Figs. 2 and 5).

The evolution of losses in the SSP3-7.0 scenario is strikingly similar to that in the SSP2-4.5, particularly for the CRH experiments. The 2040 losses are almost identical between the two scenarios, although the SSP3-7.0 result has slightly higher losses at each return period. AALs in both the mid to late century start at about 5% in the CRH result, tapering down from an early century high of about 25%. At the 500-year return period, losses increase slightly and return back to a break-even point at a 1,000-year recurrence. On the other hand, the SD experiment starts with an AAL of 10%, improving to 5% at the 100 and 200 recurrence periods and breaking even at 1,000 years. In the SD experiments, losses continue to decline at each return period over time, consistent with a declining TC frequency with warming (Fig. 2).

In assessing the outcomes from individual models, it is clear across all epochs and scenarios that both the CRH and SD experiments demonstrate a notable range in losses, a trend that persists irrespective of the moisture variable, although more pronounced in the CRH experiment (Fig. 11). In contrast, the majority of the projections show less deviation in the SD experiment, clustering more tightly around the mean losses. The most extensive range of losses corresponds to the AAL while the smallest range occurs at the highest return period, 1,000 years. The closer the epoch is to the end of the century, the wider the spread in general, as expected given increasing uncertainty in most climate parameters as time and radiative forcing both increase.

4. Discussion

4.1 Nonlinearities in the TC Response to Radiative Forcing

The relationships of the ensemble mean TC frequency in the ATL TCGI CRH experiments to net radiative forcing and global mean surface temperature are strikingly nonlinear, and are quite similar in the SSP2-4.5 and SSP3-7.0 projections, although the latter has substantially stronger forcing. In both scenarios, TC frequency decreases slightly with time, although the global mean surface temperature increases till the end of the century in these scenarios (Fig. S1). We would be tempted to conclude that TC frequency is simply insensitive to forcing in this set of models (including CHAZ CRH) but for the increase during the early century, peaking around 2040. We assume this is a forced response; why, then, does it not increase in magnitude further as the forcing does? And why does the WNP behave differently, increasing smoothly with both forcing magnitude and time within a given scenario? We cannot answer these questions confidently based on our present results, but prior work suggests two possible explanations: competition between greenhouse gas and aerosol forcings, and the role of the “pattern effect” in sea surface temperature changes.

Through its cooling effect on the ocean surface, aerosol forcing is understood to have a substantial negative impact on TC frequency and intensity (e.g., Wang et al. 2014, Ting et al. 2015, Sobel et al. 2016). Indeed, the presence of anthropogenic aerosols has emerged as a significant factor that likely contributed to the period of relatively low TC activity in the ATL from the late 1960s to the early 1990s (Mann and Emanuel 2006, Dunstone et al. 2013, Murakami 2022). Sobel et al. (2019) offer evidence that one watt per metre squared of aerosol forcing affects TC potential intensity (PI) roughly twice as strongly as the same amount of greenhouse gas forcing does, and with opposite sign, although the two have equal (and again opposite) effects on SST. This means that if greenhouse gas forcing is twice aerosol forcing, SST will increase while PI remains constant. PI largely controls simulated TC intensity in CHAZ, and when CRH is the humidity variable, PI also exerts a strong control on radiatively forced changes in frequency through its expression in the TCGI (Lee et al. 2020).

The SSP3-7.0 scenario assumes considerably stronger aerosol emissions through the end of century than does SSP2-4.5 (Riahi et al. 2018, Gidden et al. 2019, Lund et al. 2019), so greater aerosol forcing in SSP3-7.0 could be one possible explanation for the lack of increase in activity in that scenario relative to SSP2-4.5 in the ATL CRH results. In the ATL SD experiments, GHG forcing has the opposite effect (suppressing TC activity) so it should add constructively to the aerosol forcing. Figs. 2, 3 and 6 show that in the ATL, and the WNP also for that matter, the decreases in SD do become greater from SSP2-4.5 to SSP3-7.0; the changes from SSP3-7.0 to SSP5-8.5 are now small, suggesting that the increased GHG forcing compensates for decreased aerosol forcing.

Another factor that might explain some of the results, including possibly the weak dependence of TC activity on radiative forcing in the ATL CRH simulations as well as the differences between the ATL and WNP, is the pattern of projected warming in the tropical Pacific. Under GHG forcing, successive generations of CMIP models consistently project enhanced warming in the eastern equatorial Pacific compared to the western, a trend broadly similar in structure to what we observe on shorter time scales during El Niño events. It is projected that these changes will give rise to significant changes in tropical precipitation (Xie et al. 2010) and atmospheric circulation (He and Soden 2015), alter global teleconnection patterns (Zhou et al. 2014), regulate the magnitude of climate sensitivity (Andrews et al. 2015) and increase the frequency of extreme ENSO events (Cai et al. 2014). It is reasonable to expect that the effects on TC activity might also be similar to those observed in response to historical El Niño events.

With the well-documented relationship between TCs and ENSO in mind (Camargo et al. 2007, Chan 2000), we conducted some preliminary analysis, creating difference plots for yearly track density composites and box plots of TC frequency between El Niño and La Niña events (Fig. S2), covering both historical and future SSP scenarios. The models capture several characteristics of observed TC activity in both the WNP and ATL during the different phases of ENSO. During El Niño years, there are fewer landfalling storms in East Asia but more over the open ocean, while TC activity increases near Southeast Asia during La Niña years. In the Atlantic, El Niño is associated with suppressed hurricane activity, while La Niña is linked to increased hurricane activity (Figs. S2a and S2b) (Camargo et al. 2007, Chan 2000, Camargo and Sobel 2005).

While the pattern of change in the track density remains consistent across scenarios, the magnitude varies, especially in the WNP and SD experiments, aligning with basin-wide changes in landfalling TC frequency (Figs. 2 and 6). As expected, TC frequency in the CRH experiment is higher than in the SD experiment. Globally and in the WNP, TC frequency scales with radiative forcing, and the direction of change influenced by the choice of moisture variable. In contrast, in the ATL SD experiment, TC frequency decreases with increasing radiative forcing, except in the ATL CRH results, where the change in TC activity is nonlinear. These changes reinforce the known ENSO-TC relationship across scenarios, but with the warning that TC frequency during La Niña is consistently higher than during El Niño in all scenarios. Consequently, in both the WNP CRH and SD, as well as the ATL SD experiments, the track density difference between El Niño and La Niña diminishes with increasing radiative forcing (Figs. S2c and S2d).

As discussed in detail in Sobel et al. (2023) – who looked at the same CHAZ simulations as here, in addition to other data sets – this appears broadly consistent with our results in that the Atlantic is suppressed relative to the Pacific: ATL activity increases less than WNP in the CRH results, and decreases more in the SD results. Lee et al. (in preparation) provides a more extensive examination of the ENSO-TC relationship, revealing significant variability in how this relationship is represented across different climate models and geographic regions. Their analysis shows that there is a notable lack of consensus among models, particularly in the Northern Hemisphere, resulting in inconsistencies in capturing the impact of ENSO on TC activity.

Observed SST trends in the tropical Pacific over the last half century have looked very different from the projected response to GHG forcing, with the zonal SST gradient strengthening rather than weakening, and concern has been growing that this discrepancy may to some extent result from errors in the forced response of most or even all CMIP-class models, rather than just natural variability obscuring the forced signal (Seager et al. 2019, 2022; Wills et al. 2022, Lee et al. 2022). Sobel et al. (2023) discuss the ways in which this possible ensemble-wide error in tropical Pacific SST trends could lead to substantial errors in TC activity projections in all basins. This major issue in current climate science is beyond our scope to resolve here. We simply note that it is a substantial source of epistemic uncertainty in our results.

One more possible contributing factor to the uncertainty in our results is model resolution. The performance of downscaling outputs has been shown in other contexts to depend partially on the resolution of their parent climate models (e.g., Fowler et al. 2007). A split between five 100 km and seven 250 km resolution models in our dataset provides a unique opportunity to test the extent to which uncertainty in our projections may depend on the resolution of the models being downscaled. Although the ensemble mean TC frequencies are comparable (Fig. S4), the 100 km models show greater variability than the 250 km models do (Fig. S3 and Fig. S4). This is particularly pronounced in the ATL CRH results and carries through to the loss results (Fig. S5).

The increased sensitivity and variance in our projections downscaled from the higher-resolution models may be attributable to those models’ more detailed representations of small-scale processes and feedback mechanisms. It also dovetails with our comparison of CMIP6 to CMIP5, which also indicates a significant impact of resolution (e.g., Iles et al. 2020, Moreno-Chamarro et al. 2021, Kim et al. 2022). However, our 12-member ensemble, and even more so the resolution-based subsets of five and seven-member ensembles, may not be enough to characterise the full range of uncertainty here.

4.2 Bounds on Loss Results and Uncertainty

The divergent trajectories of financial loss estimates derived from the two SSP scenarios, and even more so the two choices of humidity variables in CHAZ (Fig. 11) directly reflect the corresponding divergences in TC frequency trends (Figs. 2 and 6). This, coupled with the disagreement between individual models, underscores the fundamental importance of hazard to risk, but also raises the question of whether we can say anything useful about future TC risk. Lee et al. (2023), studied a similar divergence in CHAZ results downscaled from CMIP5, and attempted to assign likelihoods to the CRH vs. SD results for the Atlantic, and found a slightly greater likelihood for the SD, but not by enough to be conclusive (nor did they consider the WNP or other basins). It is relevant here that while the sensitivity of TC frequency trends to the choice of humidity variable is peculiar to CHAZ, the same uncertainty is shared by the field as a whole, in that the sign of the global TC frequency response to GHG-forced warming remains unresolved (Sobel et al. 2021).

We also recognize that the decision to downscale only a single ensemble member from each CMIP6 model could be another source of uncertainty in the result, since a single realisation may not accurately capture the range of internal variability in any one model. Although our multi-model ensemble approach aims to mitigate such concerns by encompassing an even wider spectrum of uncertainties, utilising only 12 models might still fall short of capturing the complete range.

While the divergence in the loss projections between the TCGI CRH and TCGI SD experiments conveys an obvious and uncomfortable uncertainty, it may yet have utility in that it places bounds within which actual losses are likely to fall. Such bounds on potential losses can inform the design of financial instruments such as insurance policies and guide investments in resilience and adaptation measures. In addition, given the uncertainty in future climate conditions, strategies that are robust to a wide range of outcomes (from the lower to upper bound) can be more resilient in the face of unexpected changes. Decision-makers can, for example, prepare for the worst-case scenario (upper bound) while hoping for the best-case scenario (lower bound).

Overall, one noteworthy aspect of our results is simply that the changes in losses are not particularly dramatic, particularly on the increasing (CRH) side, with loss increases in the Atlantic mostly less than 10% across time periods and scenarios. Loss changes in other basins (not estimated here) might be larger, given that the CRH results show greater increases in TC activity in, for example, the western North Pacific than in the Atlantic, though again such potential basin-basin comparisons should be considered in light of the uncertainty around trends in the Pacific SST trends mentioned above. We also recognize that while bias correcting the CHAZ input improves the practical utility of the uncertainty bounds, it can also mask the true range of the uncertainty in models.

5. Conclusion

This study presents an analysis of projected future tropical cyclone (TC) activity and associated financial losses, using climate projections from the latest generation of climate models, CMIP6, in conjunction with the Columbia HAZard model (CHAZ), a statistical-dynamical downscaling model, and Impact Forecasting’s Atlantic Hurricane Loss Model v13.5 Beta. Our analysis focused on the SSP2-4.5, SSP3-7.0, and SSP5-8.5 scenarios.

As in previous studies using CHAZ (Lee et al. 2020, 2022; 2023), our analysis revealed a divergence in TC activity based on the choice of humidity variable, i.e., column-integrated relative humidity (CRH) or saturation deficit (SD) in the genesis component of CHAZ. While CRH leads to an increase in TC activity under future climate conditions, SD causes a decrease, despite similar trends when the two are applied in the historical period (Camargo et al. 2014).

Tropical cyclone frequency in the WNP scaled roughly linearly with increasing global mean surface temperature (decreasing in SD results, increasing in CRH), notwithstanding the choice of humidity variable, indicative of a relatively direct link with anthropogenic GHG radiative forcing. TC frequency in the ATL followed a similar trend in the SD experiments, thus a steady decline in TC frequency from the present to the end of the century, but behaved in a more complex and nonlinear way in their CRH counterparts. Depending on the scenario, projected Atlantic TC activity declined or remained relatively unchanged from the middle to the end of the 21st century in the latter, and the response in SSP3-7.0 was overall weaker than that in SSP2-4.5, notwithstanding the greater net forcing and global warming in the former. One explanation for this could be greater aerosol forcing in SSP3-7.0, given that TCs respond disproportionately to aerosol vs. GHG forcing (Ting et al. 2015; Sobel et al. 2016, 2019).

Overall, the Pacific trends towards greater activity than the Atlantic, increasing more in the CRH simulations and decreasing less in the SD simulations. We hypothesise that this could be due to the broadly “El Niño-like” pattern of SST trend projections, with the zonal SST gradient projected to decrease. This could also explain some of the nonlinearity in the Atlantic response in the CRH results since – like aerosols – it represents an influence that competes with the GHG forcing. Since there is now substantial uncertainty over whether this forced SST trend pattern is correct (Seager et al. 2019, 2022; Wills et al. 2022, Lee et al. 2022, Sobel et al. 2023), this injects another major source of uncertainty into our projections of TC activity.

Our projected financial loss changes were largely consistent with the associated changes in TC activity and inherit all their uncertainties. The most useful aspect of these results might be the bounds they place on future changes in risk since, within those bounds, the trends remain uncertain even in sign. We argue that this, at least qualitatively, is the state of the science. While efforts to reduce the epistemic uncertainties are certainly well justified, these – e.g., around TC frequency and possible biases in simulated Pacific SST trends – are long-term scientific challenges to which quick solutions may not be forthcoming. Therefore, efforts to understand these uncertainties as they are, and define adaptation strategies that accept them and deal with them rationally, are equally well justified. While this study has focused on understanding climate change effects on direct economic loss from wind hazard, the use of climate models here can have broader applications for risk assessment beyond just insurance.

Supplementary Figures

Available as a separate PDF; download here.

References

Acknowledgements:

This work was supported by Aon plc.

Declarations

Conflicts of interest: Two co-authors are employees of Aon, which funded this work

Handling Editor: Oliver Wing, Editor-in-Chief, JCRR

The Journal of Catastrophe Risk and Resilience would like to thank Oliver Wing for his role as Handling Editor throughout the peer-review process for this article. We would also like to extend our thanks to the chosen academic reviewers for sharing their expertise and time while undertaking the peer review of this article.

Received: January 29, 2024
Accepted: July 23, 2024
Published: September 4, 2024

Rights and Permissions

Access: This article is Diamond Open Access.
Licencing: Attribution 4.0 International (CC BY 4.0)
DOI: 10.63024/dpva-2pa1
Article Number: 02.01
ISSN: 3049-7604
Copyright: Copyright remains with the author, and not with the Journal of Catastrophe Risk and Resilience.

Article Citation Details

Fosu, B., et al, 2024. Assessing Future Tropical Cyclone Risk Using Downscaled CMIP6 Projections, Journal of Catastrophe Risk and Resilience, (2024). https://doi.org/10.63024/dpva-2pa1

Share this article: https://journalofcrr.com/research/02-01-Fosu-et-al