Various sources of historical hurricane loss information exist, but no consensus reference dataset. We collated the following published sources: Grinsted et al. (2019), Muller et al. (2025), National Centers for Environmental Information (2025), Weinkle et al. (2018), and the Emergency Events Database (EM-DAT; Delforge et al., 2025). These datasets differ in temporal coverage, reporting methodology, treatment of damage components, and inclusion of historical hurricane events (Fig. 1; Table 1). To maximise sample size, we combined these datasets and, for events with multiple loss estimates, these were averaged to avoid treating any dataset preferentially and help capture uncertainty. For hurricanes which made multiple landfalls, losses were averaged from the two sources – Grinsted et al. (2019) and Muller et al. (2025) – that provide per-landfall losses. Collating data in this way yielded 134 loss estimates for landfalling (including multiple landfalls) hurricanes for the period 1979–2024, of which all variables are available for 106 events (Table 1).

Economic loss estimates must be calibrated to present-day levels of damage, including adjustment factors to account for temporal changes in inflation, wealth and building density (Weinkle et al., 2018). Typically, normalisation is based on country-level adjustments and assumes building density may be represented by residential housing changes. Regional variations in these factors, temporal changes in building vulnerability or commercial building density, and the impacts of climate change may not be accounted for (Muller et al., 2025). Each dataset we used (Table 1) provides un-normalised and normalised hurricane loss estimates. However, uncertainty arises due to the differing normalisation methodologies and reference years between datasets. To ensure consistency, we used un-normalised data and applied a unified normalisation approach, based on Weinkle et al. (2018) and Muller et al. (2025), where loss estimates were adjusted using country-level inflation and real-wealth-per-housing-unit factors (both as of 2024), and a county-level housing unit density factor (Eq. 1). 1L2024=Ly⋅I2024y⋅WHn1+2024-yy⋅Hn1+2024-yy, where L is loss per hurricane (and year), y, I is inflation, W is real national wealth per housing unit, and Hn is housing unit density. I was determined using the annual implicit price deflator for gross domestic product (GDP) for the period 1979–2024 (U.S. Bureau of Economic Analysis, 2023). Hn density was determined within R34 using U.S. housing unit data (U.S. Census Bureau, 2024). W was quantified using an estimate of current-cost net stock of fixed assets and consumer durable goods (U.S. Bureau of Economic Analysis, 2025).

Hurricane track location, intensity and size information was obtained from the International Best Track Archive for Climate Stewardship (IBTrACS) v04r01 (Gahtan et al., 2024), provided by the National Hurricane Center. For each historical hurricane, we obtained 1 min sustained maximum wind speed, vmax⁡, minimum central sea-level pressure, cp, translation speed, radius of maximum wind (from the storm centre), RMW, and the outermost radii of 34, 50 and 64 kn wind speeds (from the storm centre), respectively, R34, R50 and R64. Each quantity was determined at the timestep before the storm centre crosses over land, so atmospheric fields are minimally impacted by land–surface interactions. However, IBTrACS data are incomplete (i.e., not all hazard variables are available at every timestep for every hurricane). Therefore, we supplemented IBTrACS with data from NOAA's HURDAT2 reanalysis (Landsea and Franklin, 2013) – specifically, the “U.S. Hurricane Impacts/Landfalls” table of landfall information collected by NOAA reconnaissance aircraft (Hurricane Research Division, 2025) and from Gori et al. (2023). Where data are missing in IBTrACS, HURDAT2 data were substituted, if available. Where data are available in IBTrACS and HURDAT2 at a given timestep, HURDAT2 data were prioritised.

Storm surge and rainfall cause damage through coastal and inland flooding. In this study, historical hurricane storm tide (maximum storm surge and tidal height) data were taken from the storm surge residual product (Copernicus Climate Change Service, 2022), derived using the Global Tide and Surge Model version 3.0 (Kernkamp et al., 2011; Wang et al., 2021) forced by European Centre for Medium-Range Weather Forecasts' fifth-generation reanalysis (ERA5; Hersbach et al., 2020). This provides hourly reconstructed historical storm tide height from 1950–present. Storm tide residual is calculated as the difference between the total water level and simulated storm-tide elevation, including the influence of storm surge and tide. Storm tide may be larger in the hours before or after a hurricane makes landfall, depending on antecedent tidal height. To account for this, we defined the maximum storm tide residual as the maximum along the U.S. coastline within a 1000 km radius of the hurricane's central coordinate and 24 h before and after a hurricane makes landfall. An example maximum storm tide residual for Hurricane Katrina (2005) is shown in Fig. S1 in the Supplement.

Historical hurricane-related rainfall footprints were derived from the Multi-Source Weighted-Ensemble Precipitation (MSWEP) dataset (Beck et al., 2019), which assimilates gauge observations and satellite data to reconstruct 3-hourly rainfall 1979–present. For each hurricane, we determined total accumulated rainfall, maximum 3-hourly rain rate, and maximum total rain accumulation per grid-point along the track, each within a 500 km radius of the hurricane centre (location taken from IBTrACS) at each timestep. An example rainfall accumulation footprint for Hurricane Katrina is shown in Fig. S2. This chosen radius is based on previous research (e.g., Stansfield and Reed, 2023), although a fixed radius may lead to some overestimation of cyclone-related precipitation (Stansfield et al., 2020). To complement MSWEP, the maximum rainfall accumulation along each hurricane track was collated from National Oceanographic and Atmospheric Administration (2025), providing single maximum rainfall accumulation values, although not allowing differentiation between multiple landfalls with this dataset.

This study considers only landfalling hurricanes, excluding bypassing hurricanes. It is difficult to obtain a comparable measure of hurricane intensity for a bypassing event because its intensity, measured close to the system centre, is over ocean and may therefore be relatively high compared with a directly landfalling event. It is necessary to avoid introducing such an artifact into our statistical model.

We took county-level building value information across the U.S. from the National Risk Index (Federal Emergency Management Agency, 2024b), derived from Hazus 6.1 (Federal Emergency Management Agency, 2024a), providing 2022-relative valuations per county based on the 2020 U.S. Census. Building values are time-invariant. Near present-day (2019) building value was quantified using the LitPOP dataset (Eberenz et al., 2020), providing global aggregated building value estimates at a 1 km spatial resolution. Building value estimates vary between the two datasets; hence, we used two building value datasets.

We also used decadal, county-level housing unit density data (U.S. Census Bureau, 2024), available 1950–present, as well as semi-decadal building density estimates from the Global Human Settlement Layer (GHSL), providing built-up surface area data derived from Sentinel-2 composite and Landsat satellite imagery from 1975–present at 1 km spatial resolution (Pesaresi and Politis, 2023). Annual gridded population density data between 2000 and 2020 at 1 km spatial resolution were obtained from WorldPop (2018). For hurricane landfalls outside the temporal coverage of GHSL and WorldPop, we used data for the closest available year. Therefore, a limitation to highlight is that we overestimate population density for pre-2000 events.

Structural vulnerability – the susceptibility of buildings to damage – is influenced by building age, often used as a proxy for building condition and resilience, and related to changes in building codes (regulated construction standards that include a minimal resistance to extreme weather). We used county-level average building age data (U.S. Census Bureau, 2024) and a county-level indicator of building resistance to extreme weather (Federal Emergency Management Agency, 2025). We also used county-level Hurricane Risk Score and Hurricane Historical Loss Ratio data from Federal Emergency Management Agency (2024b) and Zuzak et al. (2021). Hurricane Risk Score is defined as the average social vulnerability (i.e., extent to which specific social groups are disproportionately susceptible to hurricane impacts) and community resilience (i.e., capacity to prepare for, withstand, and recover from hurricane hazards), and is given as a percentage. Hurricane Historical Loss Ratio is defined as the percentage of a location's exposed value damaged by past hurricanes. We quantified the mean Hurricane Risk Score and Hurricane Historical Loss Ratio across all counties within several hurricane radii (R34, R50 and R64 – see Sect. 2.2). GDP data for each affected U.S. state per year, indicating a state's resilience resources, for the period 1998–2024 (U.S. Bureau of Economic Analysis, 2025) were used, and 1998 GDP values were applied to pre-1998 events.

Hurricane wind vulnerability may also be expressed as a logistic–cubic wind–damage function, relating wind speed to the fractional value of assets lost. This study uses a simple quantification of this (Eqs. 2 and 3), which was deduced by Emanuel (2011). 2f=vn31+vn3, where f is the fraction of the property value lost and vn is defined as: 3vn=v-vthresh,0vhalf-vthresh, where v is maximum wind speed, and vthresh and vhalf are the thresholds at which no asset damage and half asset damage occur, respectively. Building characteristics (e.g., construction type and age) influence vulnerability curves, and Vickery et al. (2006) and Federal Emergency Management Agency (2024a) suggest vhalf values in the range 120–160 kn. In this study, damage estimates were computed for historical hurricanes using this single vulnerability function for all buildings, with vthresh = 40 kn and vhalf = 140 kn. However, weaker systems may be damaging, such as tropical depression Allison (2001), modelling suggests vhalf may be as low as ∼ 50 kn (Federal Emergency Management Agency, 2024a). Multiple vulnerability functions for different building types cannot be applied due to incomplete localised building characteristic data. Instead, asset damage potential was applied to 1-dimensional LitPOP exposure and GHSL building density data within each hurricane footprint to estimate the exposure value and number of buildings damaged. At each timestep, vmax⁡ was used with Eqs. (2) and (3) (Emanuel, 2011) and the extracted exposure and building density, allowing vmax⁡ to vary with time.

Hurricane size (i.e., R34, R50 and R64) data are only available in IBTrACS from 2002 onwards. This is a significant constraint for studies of impact, as size information is needed to determine the region impacted by an event – i.e., its footprint. The exposure and vulnerability impact footprints derived in this study require hurricane size estimates. To estimate size prior to 2002 and extend the sample of landfalling hurricanes for this study, we developed a skilful random-forest statistical model to estimate R34, R50 and R64 for each hurricane and at each timestep, based on vmax⁡, RMW, cp and latitude (Fig. 2), with RMW found to be the most influential predictor. When R34 estimates from this model are compared with R34 observations from 2002 onwards, a Spearman's correlation coefficient, ρ, of 0.85 and a mean absolute error (MAE) of 24.5 nm were found (Fig. 2). This evaluation used a leave-one-out approach (i.e., the prediction model was trained on all observations except one, and skill evaluated on the left-out observation) across 3984 timesteps (for which R34, vmax⁡, RMW, cp and latitude within IBTrACS are not missing). There is a slight underestimation of R34 at lower values and slight overestimation at higher values, but the model overall performs well. This statistical modelling is a skilful supplement to missing IBTrACS data (Figs. 2 and S3) and allows storm size estimation back to 1979, which more than doubles our historical hurricane event sample size. Prediction models were also developed to estimate R50 and R64 for historical storms, where observations are available, with evaluation shown in Fig. S3.

From 1979 to 2002, however, there are instances where RMW data are missing from IBTrACS (vmax⁡, RMW, cp and latitude are less often missing). So, for these timesteps, we either substituted the missing RMW value from the corresponding value from HURDAT2 or obtained the RMW value from reconstructions of Gori et al. (2023). Reconstructed RMW from Gori et al. (2023) is based on the hurricane wind model of Chavas et al. (2025) and ERA5 data. Of all 3-hourly timesteps where a hurricane is over land between 1979 and 2002 (approximately 10 000 timesteps), approximately 25 % timesteps have a missing RMW estimate from each of these three datasets. In these instances, we replaced the missing value with the RMW from the previous timestep. Although this introduces uncertainty, R34, R50 and R64 estimates from the random-forest model using RMW estimates from previous timesteps are also a function of cp, vmax⁡ and latitude, quantities that are much less frequently missing in IBTrACS, and these constrain our R34 statistical model even when using substituted RMW values.

For each hurricane, exposure and vulnerability data were extracted within the R34, R50 and R64 radii at each timestep, and accumulated to create two impact footprints: (i) along the full hurricane track where vmax⁡ > 34 kn (Fig. 3) and (ii) from the immediate landfall location (i.e., up to 12 h after landfall). Hurricanes weaken over land, so obtaining two impact footprints captures the immediate landfall, where intensity influences impacts most strongly, and the full hurricane track.

We assessed the skill of three independent statistical approaches to predict hurricane loss, each based on the same set of inputs. These are: (i) a weighted combined-rank framework, (ii) a linear-regression framework, and (iii) a random-forest decision-tree framework. Our target predictand is average loss per hurricane, derived from multiple datasets (see Sect. 2.1.1). Overall, 106 hurricanes, for which all risk variables could be quantified (Table 1), were used to train our predictive model. To evaluate the skill of the linear regression and random-forest predictive models, leave-one-out cross-validation was used, where each input case was treated once as the test case and the model trained on the remaining cases. The leave-one-out approach is better suited to evaluating the skill of single predictions than, for example, k-fold cross-validation, where input data are split into training subsets. For the weighted combined-rank framework, which determines the combined loss rank from various hazard, exposure and vulnerability ranks, such validation is inappropriate, as the combined-rank approach does not require training data. Instead, optimal combinations of weights between -10 and 10 were determined for each combination of input variables, to minimise a cost function, which in this case was the model root-mean-square error (RMSE) between predicted and observed hurricane loss rank.

Two approaches were considered for representing input variables: raw values and ranked values. As input variables span disparate ranges (e.g., cp spans 900–1000 hPa; LitPOP exposure spans USD 10 million to 1 trillion), raw values were normalised to the same scale as ranked values 1–n (i.e., 1–106). In this normalisation, each variable was linearly rescaled to assign the maximum value a score of 1 and minimum a score of 106, with intermediate values mapped proportionally between these bounds. Linear and normalised input predictor ranks were derived, with rank 1 corresponding to the costliest event and rank 106 to the least costly. To predict each hurricane's loss, the equivalent predicted loss rank and observed loss rank were identified, with the loss from that observed hurricane being attributed to that predicted loss rank.

We quantified model performance with the Spearman correlation coefficient, ρ, testing the ordering of prediction according to correlation between ranked values (i.e., whether one event is more damaging than another), Pearson correlation coefficients, r, testing non-ranked correlation, and RMSE. Spearman's ρ and RMSE were used as ρ is a good indicator of the capability of accurately predicting loss rank, but not suitable for identifying cases where large differences occur in the loss prediction of a single hurricane.

Hurricane-related losses across the U.S. have generally increased over time and exhibit large interannual variability (Fig. 4). The most destructive year for losses was 2005, with USD 153 billion in un-normalised (and approximately USD 350 billion in normalised) loss. The most damaging event was Hurricane Katrina (Fig. 5), although uncertainty is evident across available loss datasets (Table 1). If Katrina occurred today, loss may be in the range of USD 190–290 billion (Fig. 5). Katrina, however, is one of numerous high-impact hurricanes whose vmax⁡-based Saffir–Simpson category at landfall is at odds with the magnitude of associated loss (Bloemendaal et al., 2021). Katrina was a category-3 landfall, despite causing unprecedented, record-breaking damage (Fig. 5). Other cases whose damage is mismatched with their Saffir–Simpson category include category-1 Hurricane Sandy (2012) and category-2 Hurricane Ike (2008), which each caused substantial losses (Fig. 5). Additionally, Hurricane Harvey (2017) was a category-4 landfall and the second-most damaging hurricane, but its loss is uncertain, with estimates between USD 90–190 billion. The historical record reveals the limitation of an event's Saffir–Simpson category in conveying its loss.

Although loss generally increases with the Saffir–Simpson landfall category and therefore wind speed (Fig. 5), the Saffir–Simpson scale has limited skill in predicting loss (Fig. 6a), with a correlation of ρ = 0.53 across 106 events (RMSE = USD 17.1 billion when using Saffir–Simpson category to predict loss rank and associated loss using a linear model). Using vmax⁡ to predict normalised hurricane loss yields a correlation of ρ = 0.52 (RMSE = USD 17.6 billion), which is slightly lower (Fig. 6b) and consistent with Klotzbach et al. (2020). Notably, landfall vmax⁡ is significantly less skilful at predicting loss for more extreme storms, with generally higher error at more extreme vmax⁡ values (Fig. 6b), which is particularly problematic as an inaccurate forecast for these more intense storms would produce larger errors in loss. The relationship between vmax⁡ and loss is potentially nonlinear but spread, and therefore error, in loss generally increases with vmax⁡. Performing a rank correlation between observed loss and loss predicted from vmax⁡ reveals significant spread (and heteroscedasticity) across the observed range (Fig. 7a), and this is found for all events as well as those where loss exceeds USD 1 billion (Fig. S4).

Although historical losses generally increase with landfall wind speed (and Saffir–Simpson category), hazard information alone has insufficient skill in predicting historical losses, with a large RMSE across all events, a finding which substantiates previous research (Klotzbach et al., 2020, 2022a). We find various hurricane-centred hazard, exposure and vulnerability quantities correlate significantly with losses (i.e., r > 0.5), but using a weighted-rank-sum approach to optimally combine these predictors yields markedly more skilful loss predictions (ρ = 0.89 for all storms). This high correlation between loss observations and predictions, with a significantly reduced RMSE of USD 7.0 billion, improves the predicted loss rank of the most impactful historical cases. This optimal prediction model was derived by determining the weighted sum of normalised ranks of landfall cp, maximum rainfall accumulation, and the percentage of total GHSL building density damage within R64 and within 12 h of landfall. This model includes two hazard attributes representing hurricane intensity (cp and rainfall), building density impacted, and applies a vulnerability function (percentage of building damage proportional to vmax⁡). For this model, important limitations are related to the observational uncertainties within the input predictor data, as well as in the target variable. There is significant variance among loss estimates for historical events. Additional uncertainty stems from the lack of complete cyclone size information prior to 2002, which is not completely mitigated by our statistical estimation approach and could impact skill for pre-2002 landfalls. A key data gap is the availability of vulnerability information and its temporal granularity. Lastly, historical loss uncertainty necessitated consideration of multiple datasets and averaging losses where required.

Studies based on physical climate model outputs provide evidence of significant interannual to decadal variability in damage (Lavender et al., 2022) and substantial inter-model uncertainty (Meiler et al., 2023). Projected increases in vulnerable assets, assuming no adaptation, may worsen damage to a greater extent than physical hazard changes (Gettelman et al., 2018). Regional vulnerability, particularly building characteristics, is important, and regional factors are accounted for in high-resolution catastrophe models (e.g., Eberenz et al., 2021) and synthetic tropical cyclone models (e.g., Meiler et al., 2022), although results exhibit sensitivity to model setup (Meiler et al., 2025). Open-source catastrophe models may underestimate historical losses (König, 2017; Welker et al., 2021), so a comprehensive intercomparison of such models across historical hurricanes is warranted. While our statistical model may not predict hurricane losses events outside our sample (i.e., loss significantly exceeding that of Hurricane Katrina), it ranks losses in historical context and complements other catastrophe modelling approaches.

We devised a “Hurricane Predictive Loss Scale” in which hurricanes are assigned a “loss category”. This characterises hurricanes according to economic losses, complementing the Saffir–Simpson scale and prior work (Bloemendaal et al., 2021; Pilkington and Mahmoud, 2016). Our model-predicted “loss categories” matched observed categories (determined from loss data) in 70 % of cases, comparing favourably with the Saffir–Simpson scale. An illustrative example is Hurricane Sandy (2012), a lower-intensity, larger-size storm that is assigned “loss category 4”, a more appropriate characterisation of the actual loss. The HPLS offers a simple method to embed information about potential losses for forecast landfalls to support decision-making. In (re-)insurance, pooling risk across a diversified portfolio reduces capital costs and increases loss predictability (Ciullo et al., 2023). For hurricanes, however, recent basin-wide (Klotzbach et al., 2022b) and near-coast (Qi et al., 2025; Wang and Toumi, 2021; Zhong et al., 2026) trends in hazards will increasingly challenge risk stakeholders. We therefore anticipate that our scheme will be a useful risk-management tool.