Asset-level assessment of climate physical risk matters for adaptation finance

Database model

We develop a database model to collect and logically connect extra financial, climate and financial information of individual physical assets and firms. The database model provides a granular and comprehensive overview of the characteristics of firms by collecting information on their productive assets, their business lines’ composition and performance, and their financial and climate characteristic. We disaggregate firm-level information by asset and geography, considering the firm as a portfolio of business lines and geographically distributed assets⁴³. This enables us to downscale climate risk assessment to the fundamental business units of the firm, considering their potentially different exposure to climate-related hazards, due to the geographical location of their productive assets.

First, we collect data on firms’ revenues by business line. We leverage information on business units, product types, and their respective sales quantities and prices. Second, we retrieve, clean and consolidate a database of physical asset exposures from different data providers, for instance, Refinitiv Eikon (https://eikon.thomsonreuters.com/index.html), and S&P (https://www.spglobal.com/en/). As different databases provide different information on assets (e.g., some databases provide a monetary value, or a location, and some do not) as well as on owners (e.g., databases using different identifiers, or no identifiers but just firms’ names), we set up a preprocessing pipeline to overcome these data fragmentation and comparability problems. Our process attaches to each asset a location, production capacity, monetary value, useful residual life, technology, operating status and ownership. We focus on energy-related or energy-intensive assets such as power plants, liquified natural gas (LNG) facilities, and mines, given the relevance of physical risk for the energy sector (see ref. ⁴⁴ for a comprehensive review) and its consideration as “sin stocks”⁴⁵. Finally, we connect assets to business lines and thus firms’ equity valuation.

Our combined databases return 123,340 physical assets globally. In total, 3493 are located in Mexico and we reconstruct their chain of ownership. In fact, asset-level datasets generally have information on direct owners, but not necessarily on the listed owners who issue financial contracts. To solve this issue, it is necessary to match owners’ names and identifiers across multiple databases. By doing so, it is possible to reconstruct the chain of ownership from the asset to its listed owners and the issued financial contracts. See also Supplementary Section 1 for a list of used databases.

Ultimately, we link 1820 physical assets to 177 firms, both Mexican and internationally owned, that own the assets and are invested in by European financial actors. To these firms, we link 17,147 individual equity holdings of 1014 European investors consolidated via 199 different equity instruments. The total exposure value of European investors amounts to 290.11 billion USD (as of June 30, 2020).

Climate physical risk assessment

Assessing physical risk on corporate securities differs from assessing physical risk for sovereign ratings or debt, especially in terms of data availability. In fact, country-level information on past disasters is available from publicly available databases (e.g., EM-DAT³⁷), but historical firm-level information is missing. Thus, we rely on a probabilistic risk assessment approach to assess asset-level and firm-level losses. In addition, our methodology includes a dedicated treatment of the transmission channels and implications of climate physical risks on firms’ business lines, on economic sectors and macroeconomic variables, the latter being captured with the use of a dedicated macroeconomic model (ICES).

Asset-level assessment

We use the CLIMate ADApt (CLIMADA) model (https://wcr.ethz.ch/research/climada.html)^{13,20,21,35,46} to perform a probabilistic assessment of damages from tropical cyclones at each asset location, for different Representative Concentration Pathways (RCP) scenarios (2.6, 4.5, 6.0) at different years (2035, 2040, 2045, 2050). Supplementary Fig. 3 illustrates the workflow’s schematic.

We consider historical data on tropical cyclones between 1950 and 2021 provided by the International Best Track Archive for Climate Stewardship (IBTrACS) (https://www.ncei.noaa.gov/products/international-best-track-archive), for the North Atlantic and Eastern North Pacific basins. Of 1555 historical events originated from either basin, 336 crossed Mexico. A map of these events is included in Supplementary Fig. 4. We standardize events’ tracks by interpolating wind speeds at half-hour time steps. Building on Aznar-Siguan and Bresch¹³, we simulate 50 synthetic tracks for each historical event, including track decay after landfall, for the probabilistic assessment. We remove duplicate hazards in the set to obtain a final dataset of 16,728 cyclones. Hazards can be duplicated in the first place as we are using two different basins as reference points, and tracks can cross from one another.

Tracks are then mapped to centroids in Mexico, i.e., geographical points where we define a wind speed from the track. The grid is set at 0.2 degrees of latitude/longitude, for a total of 14,076 centroids matched to hazards. We also tested the effect of using a finer grid on asset-level damages for the year 2040, see Supplementary Section 4 Asset-level damages with a finer grid. The comparison shows that using a finer grid could lead to slightly higher asset-level damages. We use CLIMADA to perturb the tracks for future climate change impacts for a given RCP scenario and year. The procedure followed in the model is based on the results obtained in ref. ⁴⁷ for RCP4.5. Changes in tropical cyclones’ frequencies and intensities are then obtained by linear interpolation for different RCPs and years. In this study, we use RCP2.6, 4.5 and 6.0, and years 2035, 2040, 2045 and 2050. The choice of RCPs and years is made to match the set-up of the ICES and CDDM models. For limitations of this procedure, see CLIMADA’s documentation and references therein (https://climada-python.readthedocs.io/en/stable/tutorial/climada_hazard_TropCyclone.html).

By combining the wind speed at the centroid closest to a given asset and a damage function, we obtain asset-level impacts. The damage function describes the relation between the wind speed and the damages to a given asset. The formulation used is shown in Eq. (1)⁴⁸.

where

Equation (1) enables the translation of wind speed (W_spd) into direct damages to assets described by the fraction of damaged property F_index via a cubic power. It also considers a lower bound W_thresh of no damage occurrence and a value W_half where half the damage occurs. We follow the calibration provided in ref. ³³ for Mexico and select W_thresh to be 65 km/h and W_half to be 253 km/h. Other calibrations exist in the literature, for example, ref. ⁴⁹, which sets W_thresh = 92.52 km/h and calibrates W_half with two different approaches to either 214.56 km/h or 238.68 km/h. In comparison, the calibration in ref. ³³ may overestimate the damages from low-category hurricanes and underestimate the damages from high-category hurricanes. Both the calibrations by Dunz et al.³³ and Eberenz et al.⁴⁹ are based on the shape of the damage function proposed by Emanuel⁴⁸. However, the former calibration is performed on disaster damage data from Mexico only, while the latter is performed on disaster damage data from Mexico and the Caribbean (for W_half), and on disaster damage data from the US (for W_thresh, consistently with ref. ⁴⁸). We use the calibration by Dunz et al.³³ for this study, as it is specific to Mexico only. The damage function in Eq. (1) considers only wind speed. This is a common assumption in the literature^13,33,48. Nevertheless, considering only wind speed limits the extent of the assessment of those hurricanes where rainfall and storm surges can account for high damage, despite the storm being less windy¹³. Importantly, we keep the damage function constant across asset types. Calibration of asset-level damage functions is left for further research.

We use two measures of damages at each scenario-time combination: Expected Annual Impacts (EAI) and 250 years Return Period (RP250). These combine the damage functions and the hazards to obtain measures of average (EAI) or tail (RP250) risks on assets. EAI is computed as:

where x_ij is the realization of the random variable X representing the impact, index j denotes a physical asset, E_i is an event, F its annual frequency and N_ev is the number of (independent) events considered.

For cyclones, return periods are defined as “the frequency at which a certain intensity of a hurricane can be expected within a given distance of a given location” (https://www.nhc.noaa.gov/climo). For example, a return period of 20 years for a hurricane means that on average during the previous 100 years, a hurricane of a certain category or greater passed within 50 nautical miles (58 miles) of a given location about five times. Importantly, a 1-in-100-year event will not necessarily occur once in a century but may also occur more often, or not occur.

For more details on the estimation of return periods, see ref. ⁵⁰. For the implementation in CLIMADA, the reader is referred to refs. ¹³ and ²⁰ and the model documentation (https://climada-python.readthedocs.io/en/stable/index.html).

Macroeconomic assessment

We use the ICES model (https://www.icesmodel.org/) to quantify macroeconomic impacts of chronic risks^23,24, as applied within the COACCH project (https://www.coacch.eu/)²². We source gross domestic product (GDP) and sectoral output trajectories under different combinations of scenarios (Shared Socioeconomic Pathways (SSPs) and RCPs), assumptions on capital mobility, and level of climate change impact. Trajectories are provided as “baselines”, i.e., without climate change, and as “impact scenarios”, i.e., including climate change. Thus the output change from a baseline to an impact scenario is dependent on climate change only. We use the following SSP-RCP combinations for our study, at the time horizons 2035, 2040, 2045, 2050: SSP2-RCP6.0, SSP3-RCP2.6, SSP3-RCP4.5, SSP5-RCP4.5. For the purpose of this study, we use the assumptions of high climate change impacts and high capital mobility in ICES. The ICES model as used in COACCH is resolved in 5-year steps until 2070, though for this study, we use a 2050 horizon. Further description of ICES’ sectors and the scenario choice are provided in Supplementary Sections 2 and 5.

Climate financial valuation: the Climate Dividend Discount Model

We quantify climate physical risk adjustments on equity valuation by developing a Climate Dividend Discount Model (CDDM). It extends the traditional Dividend Discount Model (DDM) framework^25,26 in its three stages formulation²⁸ to account for acute and chronic risks on firms’ long-term growth. The former depends on assets and extreme events, the latter on business lines and their economic trajectories. To estimate the market value of equity, DDMs discount the future dividends using a discount rate to determine their present value. The discount rate represents the rate of return required by investors. Alternative formulations of this discounting concept exist, for example, based on Discounted Cash Flow (DCF). Importantly, our methodology can be applied to DCFs too. We assume the following:

• Dividends can be estimated by combining Earnings Per Share (EPS), their growth, payout ratios, and their respective long-run trends.

• Long-term climate physical risks are not accounted for in the current valuation⁵¹. Thus, the long-run growth rate used for equity valuation is not consistent with future climate risks. We assume this growth rate without climate impacts is constant across all firms.

• Physical risks are going to impact mostly the long-run part of the valuation.

• Discount rate is constant for all firms in all periods.

The equity value at time 0 (V₀) is computed as:

where D_t represents the dividend at time t, r the discount rate, and \({V}_{{t}_{2}}\) the terminal value once the explicit estimation of dividends ends. t₁ and t₂ are the boundaries of the first and second stages. The following relation connects the dividends to Earnings Per Share (EPS):

where EPS_t represents the earnings per share at time t, and b_t the earnings retention rate making (1 − b_t) the payout ratio. We obtain EPS and Dividends Per Share (DPS) from S&P. Data are available for a generally limited number of years. Thus, to complete the dividend series until t₂, we first estimate EPS as:

where g_EPS,t,t−1 represents the EPS growth rate between t − 1 and t, and it holds \({g}_{EPS,{t}_{2}}={g}_{L}\), where g_L represents the long-run growth rate of dividends. Importantly, this relation implies a linear decline of EPS growth towards g_L. Thus, Eq. (6) enables the estimation of missing EPS, and hence dividends from Eq. (5). Then, the three stages in Eq. (4) can be distinguished as follows. First, a stage where dividends are explicitly estimated by a data provider (for this study, S&P). Second, a stage where dividends are modelled with linear decline. Third, the estimation of a terminal value.

We can rewrite Eq. (4) for each firm as:

where g_L represents the long-term growth rate of dividends, i.e., the rate at which the firm reaches an equilibrium where investment opportunities, on average, earn their opportunity cost of capital. The index j represents the jth firm. Differently from Eq. (4), we now make the dependence on j explicit, a necessary step for practical applications. In fact, dividend data are generally sourced from a provider whose analysts will perform the in-depth analysis necessary for dividends’ estimation. Depending on the firm, analysts will use different assumptions while modelling dividends. The usage of firm-specific assumptions for the estimation is reflected in the formula by the j subscript, which shows how dividends are going to differ from firm to firm (denoted \({D}_{t}^{\,\,j}\)) and how the boundaries of the stages are not fixed a priori but depend on the firm (denoted by t_j,1 and t_j,2). Note that dividends are explicitly modelled in the first stage, up to t_j,1 and subsequently reverted to long-run growth rates. This explicit step is necessary to link the theory of equity valuation to its practical implementation. For a discussion of the sensitivities of the model to r and g_L, please see Supplementary Section 7. For this study, we set r = 0.09 and g_L = 0.06.

Due to climate change, we cannot keep the long-run growth constant across firms as in e.g., Refinitiv’s StarMine model (https://www.lseg.com/en/data-analytics/financial-data/analytics/quantitative-analytics/starmine-intrinsic-valuation-model), as firms will be heterogeneously impacted. The growth rate of a firm will ultimately depend on: the output trajectories of its business lines as impacted by chronic risks, and the location and characteristics of its assets as impacted by acute risks. Importantly, our results shall be considered conservative due to the complexity of asset-level data, the compensation coming from positive macroeconomic shocks, and the considerations of only tropical cyclones.

We can now define the long-run growth rate as adjusted by physical risk considerations, \({\tilde{g}}_{L}\) as:

where I denotes a climate change impact scenario, O_i,B and O_i,I the output trajectories for sector i under scenarios I, impacted by climate change, and B, without climate change. We define as chronic shock the ratio \(\frac{{O}_{i,I}}{{O}_{i,B}}-1\), meaning for instance that a loss on output, relative to the baseline, of 5% corresponds to a shock of −0.05. Note that, when entering Eq. (7), \({\tilde{g}}_{L}\) is always discounted by \({(1+r)}^{{t}_{2}}\), regardless of the considered damage measure for δ_j,I,i. This implies that, when considering the RP250 scenario, the discounting for changes in g_L is still \({(1+r)}^{{t}_{2}}\) and not (1 + r)²⁵⁰. The parameter δ_j,I,i depends on the firm- and business line-specific loss due to acute risk conditioned to scenario I for sector (business line) i; s_i is the applicable revenue share for business line i, and K_j represents the total number of business lines for firm j. We design δ_j,i as a firm- and business line-specific variable computed as an average of all δ_a,j,i, i.e., the impact for all physical assets a owned by firm j contributing to business line i. In our application, δ_a,j,i depends on three main parameters for the asset: its monetary value, its residual useful life, and the impact from tropical cyclones computed in CLIMADA.

where R_a is the residual life of asset a, f(R_a) a coefficient proportional to the residual life (for the purpose of this application, \(f({R}_{a})=\frac{1}{\tau }{R}_{a}\), where τ equals 1 year), V_a is the value of asset a, L_a is the impact on asset a. Thus, η_a represents an estimate of the relative impact (i.e., fraction of the asset value) on assets from tropical cyclones, taking into account the residual life of the asset. Importantly, δ_j is floored to 1 and capped to 2, i.e., we assume that firms do not benefit from having physical assets less exposed to climate physical risk, hence simply follow the general sectoral trajectories of their business lines. This assumption is made for simplicity, since assessing the existence of positive effects stemming from asset location requires an analysis which is beyond the scope of the paper. Note that the estimate of δ_j,I,i is computed on available assets but applied to the full business line.

Combining Eqs. (7) and (8) we obtain the CDDM formulation for the adjusted equity value (\({\tilde{V}}_{0,I,j}\)):

Where \({\tilde{g}}_{L,I,j}\) is the adjusted growth rate as per Eq. (8) and \({\tilde{V}}_{0,I,j}\) is the adjusted equity valuation conditioned to a given impact scenario I. Thus the combined equity shock, stemming from the revaluation of shares considering chronic and acute physical risks, is given by:

To interpret the equity shock we proceed as follows. The asset-level impact from tropical cyclones is a random variable I_a. Characterizations of its distribution in terms of EAI and RP are provided by CLIMADA. Since I_a is a random variable, then \({\tilde{g}}_{L}\), \({\tilde{V}}_{0,j,I}\) and \( {\psi }_{j}\), being functions of I_a are also random variables and the distributions of their values could be generated from the distribution of the values of I_a.

However, this is a computationally expensive procedure. For the purpose of this paper, we extract selected moments and quantiles of the distribution of I_a, and compute \({\tilde{g}}_{L}\), \({\tilde{V}}_{0,j,I}\) and \( {\psi }_{j}\) only for those moments and quantiles. Specifically, we select the first moment, i.e., the expected annual impact (EAI) and the 99.6th quantile (RP250). Note that with regard to the estimation of Value at Risk, since the loss on equity valuation is an increasing function of the economic loss, the quantiles of the adjusted equity valuation (\({\tilde{V}}_{0,j,I}\)) are equivalent to the adjusted valuation computed on the quantiles. In contrast, regarding the estimation of expected annual impact, we approximate the average of the adjusted equity valuation (\({\tilde{V}}_{0,j,I}\)) with the adjusted valuation computed on the average of the impact (I_a). Hence, we assume that using EAI for the equity valuation leads to computing the average adjusted valuation. Similarly, we assume that using RP250 for the equity valuation leads to computing the Value at Risk (VaR) of a firm’s adjusted valuation. As such, for firms, the equity value computed using EAI represents an average equity value from the distribution of possible values considering physical risks. Similarly, the equity value computed using RP250 represents a percentile, or a VaR, of the distribution of possible equity values considering physical risks. Thus, the interpretation of the adjusted equity value is as follows. The firm has multiple possible growth paths in the long run. The equity value computed using RP250 corresponds to a future where the firm suffers damages from tropical cyclones that are comparable in magnitude to the ones emerging from an RP250 hurricane conditioned to a given climate scenario. Similarly, the equity value computed using EAI corresponds to a future where the firm suffers damages from tropical cyclones that are comparable in magnitude to the yearly expected damages. Importantly, also O_i,I is a realization of a random variable and we can interpret it to be an average chronic risk impact. Analysing the relation between the two random variables is out of the scope of this paper. As such, we treat the realization O_i,j as an average of chronic risks. We assume the equity valuation computed with the product of O_i,j and average acute risks (EAI) approximates the average equity valuation considering physical risks. Similarly, we treat the combination of the realization of O_i,I and RP250 as a (tail) percentile. The interpretation is the same also in the presence of chronic risks: firms are supposed to be exposed to effects corresponding to a given hurricane and to chronic effects as described by O_i,I, conditioned to a given climate scenario. Also, the adjusted equity value computed combining chronic risk and EAI still represents an average value, and the adjusted equity value computed combining chronic and RP250 still represents a VaR.

Importantly, in Eq. (8), δ_j,I,i is computed using the relative damages to assets as calculated by the CLIMADA model. Thus, in our model the impact of acute shocks on the firm is captured in a reduced form as adjustment in the growth rate of the firm. In this treatment, the ratio of asset damages (as computed in Eq. (1) using the damage function) links acute risks to the growth rate. Hence, it is not necessary to model the growth of assets explicitly. This approach is also consistent with the one followed in the macroeconomic model ICES. In fact, ICES represents the impacts of climate change either as changes in productivity or as losses on physical capital and land. Thus, the focus on relative losses to assets is consistent with the treatment of physical capital in the macroeconomic model.

The CDDM is computed conditioned to the following scenario combinations: SSP2-RCP6.0, SSP3-RCP2.6, SSP3-RCP4.5, SSP5-RCP4.5. In our model, the computation of the value of the firm takes into account the year span from 2022 to 2050. We assume firms are subject to climate impacts from 2035 onward, the period which captures the long run in the current treatment of the model. To proxy these impacts, we use a reference year for both the ICES model and tropical cyclones, namely 2040 (climate models’ estimates at years 2035, 2040, 2045, 2050 are not distinguishable anyway in statistical sense⁴⁷). In particular, the estimate of tropical cyclones’ impacts at 2040 is obtained following a common approach in the literature based on a linear approach interpolation of the impacts between 2020 and 2100 (see e.g., refs. ²¹ and ³⁶). Thus, for the valuation conditioned to e.g., scenario SSP2-RPC6.0, year 2040, ICES data are considered for SSP2-RCP6.0, year 2040, and tropical cyclones impacts are considered for RCP6.0, year 2040. Other years are not considered. Finally using a certain year to compute g_L does not imply extending the dividend stream until that year, but only computing Eq. (8) with values for O_j,I and δ_j,I,i for that year.

The full CDDM is applied to all firms with at least two datapoints available for EPS. For non-dividend paying stocks or stocks with missing data, we compute the equity shock as follows:

• For stocks paying no dividend, or with no dividend data, we revert to direct shocks, i.e., \({\psi }_{j}={\tilde{g}}_{L,I,j}-{g}_{L}\).

• For stocks with data available only for the first period, we compute a one period version of the CDDM following \({\tilde{V}}_{0,j}=\frac{{D}_{0}}{r-{\tilde{g}}_{L,I,j}}\).

Analysis using proxy data

We compare the portfolio-level results for losses computed using proxy data vs asset-level data. The purpose of this analysis is to quantify the relevance of the underestimation of physical risk that stems from neglecting asset-level data. To compute the results with proxy data, we replace asset-level data with one location per firm and use information on physical risk at this location to proxy physical risks for the firm. Thus, the CDDM model is computed for the same firms both using asset-level data and using only one location (i.e., proxy data). Business lines are used for comparability, i.e., the valuation is applied only to those business lines that are adjusted in the asset-level analysis. The single location is either the headquarter (for Mexican firms), the address of the Mexican subsidiary with highest ownership (for non-Mexican firms with Mexican subsidiaries) or Mexico City (for non-Mexican firms without Mexican subsidiaries). All addresses are geolocalised using Opencage API (https://opencagedata.com/) and checked manually.

For the one location, the value exposed to acute physical risks is given either by the firm’s Property, plant and equipment (PPE) (Variable “property plant & equipment—net total”, sourced from Refinitiv Eikon) (if larger than 1 million USD) or by its Total assets (TA) (Variable “total assets”, sourced from Refinitiv Eikon). Using the methodology described in Subsection Asset-level assessment, we assess impacts from CLIMADA on PPE or TA at the single location. We combine the losses from tropical cyclones with chronic risks computed from ICES and plug them in the CDDM. For comparability purposes, we use δ only on those business lines which can be impacted at the asset level. Otherwise, applying δ to business lines that cannot be impacted would impair the comparability of the counterfactual analysis. Subsequent statistics (e.g., VaR) are computed as in the version of the model using all asset information.

Limitations

The following remarks complement the limitations acknowledged in the Discussion. First, we account for uncertainties on financial portfolio loss (VaR, mean) estimating the confidence intervals (CI) using bootstrapping (see e.g., ref. ³⁰). Including the sources of uncertainties mentioned in the Discussion would likely lead to larger CI. Second, the reader should be aware that there is considerable uncertainty regarding the effects of climate change on the frequency and intensity of tropical cyclones for the middle of the century especially at less-than continental scale. The methodology we applied here to quantify damages around the middle of the century builds on relevant literature in the field and relies on interpolation (e.g., refs. ²¹ and ³⁶). Third, we do not consider firms’ adaptation measures (such as sea barriers or mangroves) due to lack of data on firms’ adaptation strategies³⁸. Adaptation may vary across firms, and different firms may follow different schedules to implement adaptation measures. Moreover, many adaptation measures generally considered in the literature (e.g., mangroves for coastal protection, as in ref. ³⁶) are not relevant for the types of assets that we analyse here. Similarly, relocation is not feasible for most of the assets we consider (e.g., mines or power plants must be located where natural resources are located). Existing calibrations of adaptation measures are either based on assumptions (e.g., ref. ³⁶ in the case of the effect of mangroves on tropical cyclones’ winds) or specific to individual countries and thus not applicable to Mexico (e.g., ref. ²¹). Furthermore, Mexico invests very little in adaptation^52,53. Fourth, information on assets’ location, ownership, value and residual life is often missing and has to be estimated. Moreover, some non-core firms’ assets (e.g., deposits, warehouses) may be unknown even for firms where asset-level data are available. Furthermore, for some assets it is not possible to reconstruct ownership chains, or the unlisted nature of some of the owners makes the link to equity financial portfolios not possible. Fifth, in our assessment, we consider only one country (Mexico), one hazard (tropical cyclones), a selection of asset types (mostly energy-related), and one financial asset class (equities). Thus, our results in terms of financial risk for investors are conservative. Sixth, short-term risks are generally downplayed both in the macroeconomic model framework used (see ref. ²³ for a discussion), and in the CDDM. Finally, we consider only direct impacts of tropical cyclones, and not their indirect ones such as supply chain disruptions, damage to infrastructure other than the assets in the sample, or loss of lives.