The hard part is assigning a parent company's balance sheet to individual facilities. That judgment carries the USD 8.2 billion estimate, and the authors explicitly say they cannot validate it.
Reconstructing the figure is chiefly a data procurement exercise. The equations take one line. The necessary inputs require several paid subscriptions plus a hydrology archive.
We ran no independent test. The method needs geocoded facility footprints and flood projections rather than a price signal. Substituting US-equity data would require building the hazard and facility inputs from scratch.
Mandatory climate disclosure creates the demand for this number. The TCFD recommendations ask listed companies to quantify climate-related financial risk across geographically dispersed operations. Since 2023, the ISSB's IFRS S2 has imposed the same expectation and incorporated those recommendations in full. Lenders with corporate loan books want a facility-level measure of flood exposure they can benchmark. Pension and insurance portfolios holding equity and bonds want one too, as do underwriters pricing claims. Abe and Adriaens make a specific novelty claim: no prior method has measured these risks by industry sector across an entire listed index.
From flood depth to corporate loss
The model has three layers. It covers all 225 constituents as of December 2024 and more than 18,000 facilities. Ownership records from Orbis, Bloomberg, EDINET and annual reports supply the locations, which are geocoded through the Google Maps API. Each site is then matched to WRI Aqueduct Floods depth grids.
Those grids span five GCMs, RCP 4.5 and 8.5, and horizons of 2030, 2050 and 2080. They also cover nine return periods from 2 to 1,000 years. The analysis uses riverine flooding, although the database, as described by the authors, includes coastal flooding as well.
Losses enter through two channels. For property damage, a depth-dependent damage ratio from the JRC industrial curves in Huizinga et al. is applied to facility PP&E. Business interruption uses a depth-dependent estimate of disruption days, multiplied by daily facility sales. Those days come from a 2023 survey by the Japanese Ministry of Land, Infrastructure, Transport and Tourism (MLIT). Expected annual damage follows from integrating loss over exceedance probability across the nine return periods. The corporate financial inputs are decadal averages from 2014 to 2023, covering 2,250 company-years of PP&E and sales.
By 2030 under RCP 8.5, expected annual damage reaches USD 8,209 million a year. Property damage contributes USD 2,852 million, while business interruption contributes USD 5,357 million. Asia accounts for USD 5,273 million (64%), followed by Europe at USD 1,289 million (16%), North America at USD 804 million (10%) and Oceania at USD 681 million (8%). Central and South America and Africa together contribute USD 162 million.
Exposure is concentrated. Among the 18,000-plus facilities, just over 2,900 (16%) show any modeled flood exposure. When losses are normalized by five-year average net income from 2019 to 2023, property damage alone averages roughly 5% of earnings for utilities. The paper says the figure "well exceeds this value when BIL is included", using BIL for business interruption loss. The supporting chart is left untabulated.
A P&L never enters the analysis. The abstract states that flood risk was not weighted by index component allocation. The limitations also instruct readers that the forecasts "should be interpreted as order of magnitude values and patterns in space and time." Quibbling over headline precision would therefore miss the dispute. The useful questions concern which judgments move the estimate and in which direction.
The imputation carries both loss channels
Private subsidiaries do not disclose sales and PP&E at facility level. The authors allocate corporate totals through linear scaling based on employee counts and addresses. They call the resulting facility figures proxies and state that this uncertainty "was not explicitly quantified for this article".
Both loss calculations depend on that allocation. It determines the PP&E multiplied by the damage ratio and the daily sales multiplied by downtime. The reported ranges deserve careful separation from this issue. Section 5.2 varies the MLIT survey and the Huizinga fixed asset damage curves in a sensitivity analysis of the flood damage functions. Its dominant parameters differ across sectors and companies. The error bars in Fig. 4 show variation across the five GCMs. Neither analysis covers the imputation. An implementer using another headcount-to-revenue allocation rule will produce another USD 8.2 billion, outside any gap bounded by the paper's reported ranges.
The downtime curve is the next major judgment. Business interruption supplies 65% of total loss, with every facility in every region valued from the same Japanese survey. Abe and Adriaens identify this limitation and cite Sultana et al. Their study of the 2013 German floods indicates longer interruption periods than the MLIT survey, suggesting that the estimated losses may be low.
Two more stated assumptions lean in the same direction. Spatial dependence among flood events is outside scope, which the authors say "can lead to underestimates of both local and aggregate corporate risk." Property damage and business interruption are also modeled independently, an assumption they acknowledge is false. A damaged plant cannot immediately return to full production, while supply links can stop operations at sites that never flood.
Other assumptions pull against each other. Financials remain fixed at their 2014 to 2023 averages through 2080, without growth or a discount rate. Companies that expand sales or assets will therefore have lower future EAD in the model than under a growing base. Meanwhile, the model includes no flood mitigation or defence between 2023 and 2080, which raises the estimate. The authors leave those effects separate, writing that "the inclusion of corporate growth rates and discount rates is likely to increase uncertainty in the assessment of the implications of future fluctuations in flood risk."
How much does the 5% threshold say?
The sector chart uses 5% of net income as its Major Risk line. Abe and Adriaens explain where that threshold comes from: they find no academic literature for choosing materiality metrics and instead cite corporate risk white papers. Those sources set different standards. One scale classifies 1 to 5% as moderate, 5 to 10% as major and above 10% as critical. Snam uses 1.5% of net profit. KPMG presents 5% as a historical rule of thumb.
At the paper's 5% threshold, several companies cross the benchmark in materials, consumer discretionary, consumer staples, industrials and information technology. Move the threshold to 10%, and average utility property damage stays below it.
The denominator changes the interpretation. Only 188 of 225 companies (84%) had positive five-year average earnings and were included in the normalization, so the sector percentages apply to the profitable subset. Utilities, the only sector whose average crosses the line, contains five companies. Their average PP&E is USD 34.63 billion. Fixed asset turnover, defined as sales divided by PP&E, is 1.02 versus an all-industry mean of 4.24. The sector result can be traced through five companies.
Country rankings have the same small-sample problem. Sudan leads on EAD per facility with one at-risk asset. France comes second with six, while Japan has 873. The authors explicitly make this point about Sudan.
Europe's unusual split
Europe contains 10% of facilities, yet it produces 22% of interruption EAD and only 5% of property EAD. Interruption is about 8.7 to one relative to property loss, at USD 1,156 million versus USD 133 million.
The paper identifies part of the reason. At 1 m depth, the JRC damage ratio is 0.27 for Europe and 0.48 for Asia. Higher fixed asset turnover in sectors weighted toward Europe pushes in the same direction. Less PP&E per unit of sales leaves fewer assets to damage and more revenue to interrupt.
The paper also proposes differences in labor laws and welfare systems as a possible influence on restart times. By construction, that effect cannot appear in the estimates because every region uses the same Japanese depth-to-days curve.
What the output can support
Imputation leaves the broad pattern intact and the level uncertain.
Interruption running at roughly twice property damage places the exposure in revenue continuity. That calls for a different insurance product and a different capex discussion from physical asset hardening. Concentration matters more than breadth. Only 16% of facilities have exposure, yet average utility losses exceed 5% of earnings. Individual companies in five other sectors pass 5% and even 10%.
Regional movement also varies. Oceania's modeled EAD declines 4% from 2030 to 2080 while most regions increase. Portfolio climate screens that assume a uniform global rise would miss that result.
Portability depends on assembling the data. Whatever the abstract says about industry-sector novelty at index scale, building the dataset is the substantive workload. A US-equity version would need fresh facility-to-parent ownership mappings for tens of thousands of sites, followed by Aqueduct depth grids and regional damage curves.
Validation remains the evidence I would want next. In the authors' 2025 paper, the equation was checked against disclosed losses from four Japanese companies. The comparison with the 2011 Thailand flood is described here as not quantitatively analyzed. Modeled facility EAD matched against realized claims across a broad sample of flood events would turn the imputation dispute into an empirical question.
That standard is mine rather than the authors'. Until it is met, the method works as an exposure screen for adaptation planning and concentration review. Abe and Adriaens go further, suggesting that EAD "can provide an indication of annual flood insurance premiums" and that the sector benchmark could assist the reinsurance industry in pricing cover.