When does a disaster become a systemic event? Estimating indirect economic losses from natural disasters
Sebastian Poledna, Stefan Hochrainer-Stigler, Michael Gregor Miess, Peter Klimek, Stefan Schmelzer, Johannes Sorger, Elena Shchekinova, Elena Rovenskaya, JoAnne Linnerooth-Bayer, Ulf Dieckmann, Stefan Thurner
WWhen does a disaster become a systemic event?Estimating indirect economic losses from natural disasters
Sebastian Poledna , , , Stefan Hochrainer-Stigler , Michael Gregor Miess , , , ,Peter Klimek , , Stefan Schmelzer , , Johannes Sorger , Elena Shchekinova , ElenaRovenskaya , , , Joanne Linnerooth-Bayer , Ulf Dieckmann , , and Stefan Thurner , , , International Institute for Applied Systems Analysis, Schlossplatz 1, 2361 Laxenburg, Austria Complexity Science Hub Vienna, Josefst¨adter Straße 39, 1080 Vienna, Austria Section for Science of Complex Systems, Medical University of Vienna, Spitalgasse 23, 1090 Vienna, Austria Santa Fe Institute, 1399 Hyde Park Road, Santa Fe, NM 87501, USA Institute for Advanced Studies, Josefst¨adter Straße 39, 1080 Vienna, Austria Faculty of Computational Mathematics and Cybernetics, Lomonosov Moscow State University (MSU), Moscow, Russia and Vienna University of Economics and Business, Welthandelsplatz 1, 1020 Vienna, Austria
Reliable estimates of indirect economic losses arising from natural disasters are currently out ofscientific reach. To address this problem, we propose a novel approach that combines a probabilisticphysical damage catastrophe model with a new generation of macroeconomic agent-based models(ABMs). The ABM moves beyond the state of the art by exploiting large data sets from detailednational accounts, census data, and business information, etc., to simulate interactions of millionsof agents representing each natural person or legal entity in a national economy. The catastrophemodel introduces a copula approach to assess flood losses, considering spatial dependencies of theflood hazard. These loss estimates are used in a damage scenario generator that provides input forthe ABM, which then estimates indirect economic losses due to the event. For the first time, we areable to link environmental and economic processes in a computer simulation at this level of detail.We show that moderate disasters induce comparably small but positive short- to medium-term, andnegative long-term economic impacts. Large-scale events, however, trigger a pronounced negativeeconomic response immediately after the event and in the long term, while exhibiting a temporaryshort- to medium-term economic boost. We identify winners and losers in different economic sectors,including the fiscal consequences for the government. We quantify the critical disaster size beyondwhich the resilience of an economy to rebuild reaches its limits. Our results might be relevant forthe management of the consequences of systemic events due to climate change and other disasters.
Keywords: resilience | large-scale data-driven modeling | economic simulator | natural hazard modeling | environmental-economic coupling | I. INTRODUCTION
Total economic losses from natural and man-made dis-asters in 2017 are estimated to be USD 306 billion. Withhurricanes Harvey, Irma and Maria, which have made2017 the second costliest hurricane season on record, theUS was hit particularly hard. Due to climate change,economic losses from extreme events – such as floods,droughts, and other climatic disasters – will further in-crease, and events that have been considered rare untilnow will become more common in the future [1]. How-ever, the resulting economic losses to a national econ-omy are difficult to quantify. While direct losses due tothe immediate destruction of homes, firms, infrastruc-tures and lives can be reliably estimated with existingdata, estimates of indirect losses – which arise as eco-nomic consequences of physical destruction – are muchharder to obtain, and no consensus on their validity exists[2]. The primary reason for this is that natural disastersaffect the economy in multiple ways and along several dimensions [3, 4]. For example, a flood event might de-stroy or damage the physical capital of a firm, causingoutput losses and worker layoffs. This, in turn, can trig-ger further output losses to suppliers and customers ofthat firm, potentially leading to more layoffs. The asso-ciated reduction in household income additionally lowersconsumption, potentially further enhancing output lossesand increasing layoffs. Subsequently, however, the recon-struction of damaged or destroyed capital can have theopposite effect: companies involved in reconstruction ac-tivities experience growth and expand their workforce,which results in higher income that ripples through theeconomy and leads to economic growth via Keynesianmultiplier effects. Hence, natural disasters simultane-ously cause both indirect losses and indirect gains. Werefer to these losses and gains as indirect economic ef-fects. To avoid double counting, we measure losses or gains as thechange in gross domestic product (GDP) relative to a baselinescenario. This definition differs from [5], who additionally at-tributes employment losses (e.g. due to the closure of damagedfacilities) to direct losses, and defines indirect losses as all eco-nomic consequences except for damages (direct losses) and em-ployment losses caused by the disaster. a r X i v : . [ q -f i n . E C ] J a n Indirect economic effects of average natural disasters,as measured by changes in GDP, are typically small. Neteffects may even be close to zero in the short term, wherelosses and gains of natural disasters can cancel each otherout [6]. Still, there are winners and losers, since effectsmay differ substantially across industries and economicsectors, and almost never cancel for companies and indi-viduals [7]. Several empirical studies even find small butpositive short- to medium-term overall effects for moder-ate natural disasters of certain types, especially for mod-erate floods [7–14]. However, the evidence on the signand magnitude of indirect economic effects of naturaldisasters in general remains mixed and conflicting, see[11–13] for recent surveys. This situation is echoed inmeta-analyses such as [15], which reports that disasterson average have an insignificant impact in terms of in-direct costs, and [16], which finds some evidence that apart of the negative impact of natural disasters reportedin studies is caused by a publication bias. Moreover,[17] shows that the impacts of natural disasters on differ-ent components of GDP (such as investment, governmentand private consumption, exports, imports) differ widelyin timing, direction and extent. More aggregate analy-ses might mask these differences, and it may be difficultto find clear and large net aggregate impacts on GDP.Severe natural disasters (systemic events), on the otherhand, are not believed to have neutral or positive eco-nomic effects on an aggregate level [7–9, 12, 18, 19]. In-direct economic losses from these systemic events may beamplified by several mechanisms such as post-disaster in-flation, network effects like supply chains affecting firmsthat were not initially impacted by the event, as wellas physical and financial resource constraints regardingthe productive capacity of the economy [3]. Despitethese general arguments, empirical evidence on indirecteconomic effects of systemic events, particularly regard-ing severe floods in developed economies, is scarce andlargely inconclusive [7–10, 20, 21].Indirect economic effects from natural disasters are dif-ficult to model with traditional approaches such as input-output (IO) models [22–28], computable general equilib-rium (CGE) models [29–34], and econometric analyses[7–19, 35] because of the over-simplifying nature of theseapproaches [36]. IO models have a tendency to overes-timate indirect economic losses and gains [2]. This wasdemonstrated, for example, for the case of job gains fromreconstruction following the Northridge earthquake inCalifornia in 1994: [37] showed that the IO estimates by[38] significantly exceeded actual data for the Los Angelesarea following the earthquake [2]. Furthermore, IO mod-els cannot incorporate the reactions of economic agents toa disaster. By design, CGE models are overly optimisticregarding the flexibility of an economy to react to naturaldisasters [3]. This is due to the underlying assumptionthat price clearing mechanisms bring the economy backto a general equilibrium after a certain time and are sub-ject only to modest constraints such as adjustment costsin the investment function. However, real-world price formation is sticky and imperfect. Production functionstypically assumed in CGE models often overstate theflexibility of substitution between factors of production(labor, capital, material input). Econometric analysis ofindirect economic effects faces the problem that statis-tical, historical relationships used to derive model pa-rameters are likely to be disrupted by the disaster [2].The latter study points out an array of factors that maychallenge the implicit socio-economic assumptions of theapproach, including: the temporary nature of measurestaken directly after the disaster event, permanent eco-nomic changes (e.g. in the production function), pur-chase and sale patterns, as well as labor force migrationor overtime hours in the reconstruction phase. Indicativeof this, [6] demonstrates that the regional econometricmodel results in [39] over-estimated the economic impactof Hurricane Andrew, which hit the US in 1992, by 70-85% [2].In summary, traditional approaches so far have failedto provide unambiguous conclusions about the indirecteconomic effects of natural disasters, in particular re-garding systemic events. This means that, given thecharacteristics of these traditional approaches, today itis still largely impossible to relate the size of initial dam-ages of natural disasters to the expected subsequent indi-rect economic effects, or to clearly disentangle expectedlosses and growth effects on a sectoral level. To addressthis problem, we propose a novel approach that combinesa probabilistic physical damage catastrophe model witha new generation of macroeconomic agent-based models(ABMs). The ABM we adopt moves beyond the stateof the art by exploiting large data sets from detailed na-tional accounts, census data, and business information,etc., to simulate the interactions of millions of agents rep-resenting each natural person or legal entity, such as cor-porations, government entities and institutions in a na-tional economy. It has been shown that this model is ableto forecast numerous macroeconomic variables includingmajor variables such as GDP, inflation, consumption andinvestment better than standard forecasting approaches[40].We apply our method to estimate indirect economiclosses from flood events in Austria, for which the men-tioned data is available in highly detailed, complete andconsistent form (see SI ). The ABM is geolocalized in thesense that the location of capital stock, e.g. firms andinfrastructure, is known. The geospatial distribution ofcapital across Austria is shown in Fig. 1. The physi-cal damage catastrophe model (damage scenario genera-tor) allows us to generate realistic virtual natural disasterevents (floods) of controlled size. It is based on a copulaapproach that assesses realistic hazard frequency and in-tensity, and takes into account spatial dependencies and Constant elasticity of substitution (CES) functions. FIG. 1. Geospatial distribution of capital across Austria. Af-fected capital in flooding zones of a 250-year event is shownin a color scale from green to red, where green indicates smalldamages and red large damages up to 10 bln. Euro. Grayindicates capital stock not affected by the flood. For exam-ple, capital stock in Vienna (the large bar in the northeast) isnot affected due to extensive flood protection measures, andis therefore shown as grayed out. the dependency on the severity of the events (see SI). Af-ter generating a geolocalized flood event of a given sizeat the beginning of year 2014, the affected dwellings,firms and infrastructure are destroyed (the affected cap-ital in flooding zones of a 250-year event is depicted inFig. 1). On this basis, the detailed indirect economiceffects across economic sectors and industries are studiedin the ABM over several consecutive years. RESULTSModerate disasters do not always have a negativeimpact on economic growth
Fig. 2 shows the indirect economic effects resultingfrom a 100-year (blue line) and a 250-year (red line) floodevent that destroys dwellings and productive capital. The total direct losses (damages) amount to about 0.6% (100-year event) and 1.2 % (250-year event) of Aus-trian capital stock, respectively. Fig. 2(a) shows the cu-mulative change in the GDP growth rate relative to thebaseline scenario in percentage points (pp). The qual- This year is chosen to simulate the flooding event since 2013 itis the last year for which the main data source – the symmetricIO tables for the Austrian economy – is available to calibrate themodel. The event occurs at the beginning of the year 2014. The baseline scenario describes a continuation of current trendsfor the Austrian economy. The baseline scenario serves as thebenchmark against which we evaluate the indirect economic ef-fects of the different flooding scenarios. A percentage point (pp) is the unit for the arithmetic differenceof two percentages. For example, moving up from 10% to 12% is
Year -6-5-4-3-2-101234 C u m u l a t i v e c hange i n G D P - g r o w t h r a t e [ pp ] (a) Baseline100-year event250-year event1500-year event
Year -3-2.5-2-1.5-1-0.500.511.52 C u m u l a t i v e c hange i n une m p l o y m en t r a t e [ pp ] (b) Year -4-202468101214 C hange i n go v e r n m en t deb t - t o - G D P r a t i o [ pp ] (c) FIG. 2. Indirect economic gains and losses of a 100- (blue),250- (red) and 1,500- (black) year flood event. Time labels onthe x-axis indicate the end of each year, and the gray verti-cal bar marks the first year after the flood. The panels showthe effects as percentage point changes relative to the base-line scenario, in which no disaster happens. (a) Cumulativechanges in GDP-growth rates. (b) Cumulative changes in theunemployment rate. (c) Changes in the government debt-to-GDP ratio. Shaded areas cover one standard deviation aboveand below the mean values, as obtained from 50 independentMonte-Carlo simulations. itative behavior of the two scenarios is similar: startingfrom small negative effects immediately after the disaster,cumulative effects on economic growth then become pos-itive in the short to medium term (2014-2018), and turnnegative in the long term. These effects are most pro-nounced with an about 0.5 pp cumulative GDP growthrate increase (250-year event) relative to the baseline sce-nario in the second year after the flood (2015). In the longterm, primarily due to a multiplier-accelerator mecha-nism [41] (see below), the effects decline to an almostneutral impact (100-year event), and to a reduction ofGDP growth of approx. 0.8 pp (250-year event), respec-tively. This behavior, i.e. positive short- to medium-termand negative long-term effects of moderate size, is in linewith the literature [7, 8, 10, 14, 35]. Fig. 2(b) demon-strates that – as to be expected according to Okun’s law– the change in the unemployment rate is inversely cor-related to economic growth, but at a slightly lower am-plitude: for the 250-year event, a cumulative decline ofalmost 0.5 pp two years after the flood (2015) is followedby a cumulative growth in the unemployment rate up to amaximum of about 0.5 pp in the long term. Fig. 2(c) de-picts the government debt-to-GDP ratio and shows thatthe dynamics of the growth and unemployment rates, aswell as the transfer we assume to be provided by the gov-ernment to fully compensate households for their lossesof dwellings as catastrophe relief, all lead to an initialrise in this ratio of about 2 pp. For three years afterthe flood (2015-2018), the government debt-to-GDP ra-tio temporarily falls slightly below its initial level, but inthe long run stabilizes at an increase by more than 2 pprelative to the baseline scenario (250-year event).
Severe disasters have pronouncedly negativeeconomic effects immediately after the event and inthe long term
A severe-disaster scenario that simulates a 1500-yearflood event is shown in Fig. 2 (black lines). The to-tal direct losses correspond to approximately 10 % ofthe capital stock in Austria. The indirect economic ef-fects after this shock are qualitatively different from themoderate-disaster scenarios. The initial overall effect onGDP growth is pronouncedly negative, with a cumula-tive reduction of GDP growth by about 5 pp, see Fig.2(a). Due to reconstruction, growth picks up fast in theyear after the disaster, and surpasses cumulative GDPgrowth of the baseline scenario by the second year afterthe flood (2015), culminating in a temporary economicboost of about 2 pp of additional cumulative GDP growthin 2016. The multiplier-accelerator mechanism [41], aswell as production, capacity and credit constraints (seeSI) drag growth downwards after this point, leading to a 2 pp increase, but it is a 20 percent increase in what is beingmeasured.
Year -5-4-3-2-1012345678 C u m u l a t i v e c hange i n G D P - g r o w t h r a t e b y s e c t o r [ pp ] Real estateConstructionManufacturing
FIG. 3. Cumulative growth effects on sectoral GDP aftera 250-year flood event for selected economic sectors in per-centage points (pp) relative to the baseline scenario. Shadedareas cover one standard deviation above and below the meanvalues. Sectors shown: construction sector (black), manufac-turing sector (red) and real estate sector (blue). The grayvertical bar indicates the year of the flood. negative long-term cumulative growth effects of approx.1.7 pp. The unemployment rate reacts strongly to the se-vere disaster, with a cumulative initial increase of morethan 1.5 pp, and is followed by a reduction up to al-most 2 pp in 2015 during the reconstruction phase, seeFig. 2(b). After this pronounced disruption of the labormarket, the unemployment rate rises again by a cumula-tive change of approx. 0.5 pp (2020) due to the cyclicaldynamics, and stabilizes at a level close to the baselinescenario in the long term. Immediately after the disas-ter, a large initial government transfer to households tocompensate for their losses of housing stock, as well assubstantial decreases in government revenues and GDP,lead to a 10 pp rise of the government debt-to-GDP ra-tio, see Fig. 2(c). Even though this ratio shortly returnsto its initial level due to the positive economic effects ofreconstruction, the downturn because of over-productionthree years after the flood event implies a subsequent risein this ratio by almost 10 pp, leaving government financessubstantially deteriorated in the long term. We assume – in line with past experiences of political processesregarding catastrophe relief by the Austrian government – thistransfer to be limited to about a third of the total losses indwelling stock.
Effects differ substantially across industries andeconomic sectors
While moderate flood events can have positive aggre-gate effects in the medium term, impacts are expectedto differ significantly across economic sectors. Fig. 3confirms this conjecture. It shows the effects for themost severely impacted sectors as a result of the 250-yearevent. The real estate sector (blue line) suffers substan-tially from the destruction of residential capital stock:sectoral output is reduced by more than 4 pp. Despitereconstruction works improving the initial situation, thecumulative growth change in this sector remains nega-tive, with a cumulative loss in growth of about 2 pp inthe long run. The construction sector (black line) imme-diately profits from the reconstruction of dwellings andproductive capital with a sectoral GDP growth of almost6 pp in the first year after the flood (2014). After thefast ramp-up of reconstruction during the first years af-ter the flood, peaking in an about 6.5 pp increase in thesecond year after the flood (2015), this effect graduallywears off in the following years, turns slightly negative byyear seven after the flood (2020) and remains rather sta-ble at this level thereafter. The restoration of productivecapital takes more time. The largest cumulative increasefor the manufacturing industry (red line) of about 1 ppis reached in year two after the flood (2015), since thissector supplies a major part of the material input forthe re-installment of losses in productive capital. Follow-ing the general downturn due to over-production and thethereby induced economic cycle, we see that cumulativeGDP growth of this sector in the long term (2019-2023)is lower by almost 1 pp than in the baseline scenario. Theeffects on all sectors can be found in Table S7 in the SI.
Loss of resilience – when disasters become systemicevents
Fig. 4 shows the cumulative changes in GDP growth(relative to the baseline scenario) as a function of thesizes of direct damages of the event for different timesafter the disaster: one year (2014), two years (2015), andthree years (2016) after the flood event. As can be ex-pected, the larger the direct damage, the larger the de-cline of GDP growth immediately after the flooding dis-aster (2014) is. In the second year after the flood event(2015), the effects of increased economic activity inducepositive overall GDP growth. Within this second year,a large part of the initial damage can already by com-pensated, i.e. cumulative growth effects in the secondyear are slightly positive, but remain below 1 % for dis-asters smaller than a 250-year event. For larger events,the economy shows a remarkable growth that is inducedby reconstruction after the disaster, which clearly out-weighs the direct losses. This growth is limited by dif-ferent constraining factors (see SI) and starts to declinewith respect to the direct losses inflicted by the disas-
Direct losses in percentage of capital stock [%] -6-5-4-3-2-1012345678 C u m u l a t i v e c hange i n G D P - g r o w t h r a t e [ pp ] Inflection point w. r. t.economic growth250-yearevent100-yearevent 1,500-yeareventMaximum: threshold atwhich natural disasterbecomes a systemic event
FIG. 4. Cumulative changes in GDP growth relative to thebaseline scenario as a function of the direct damage as a per-centage of GDP. Results are shown for three different yearsafter the disaster: 2014, 2015 and 2016. Shaded areas coverone standard deviation above and below the mean values.Immediately after the event (2014), all disaster sizes are asso-ciated with negative growth relative to the baseline scenario.In contrast, for the years 2015 and 2016 there exist inflectionpoints and maxima for GDP growth, indicating the existenceof direct damage sizes that, respectively, are “optimal” interms of economic growth and determine a threshold wherenatural disasters become systemic events. ter: the inflection point of the curve occurs at about a2.4 % loss of capital stock. The cumulative changes inGDP growth in this year show a maximum in the re-gion where the initial damage causes about 3-4 % directlosses in the capital stock. At this maximum, the growtheffects lose momentum and indirect losses start to dom-inate beyond the maximum. The situation is similar fortwo years after the event (black line): while the posi-tion of the inflection point remains at about 2.4 %, thegrowth stimulus is somewhat more pronounced, and themaximum is located at approximately 5 % direct losses.At this point economic growth is severely restrained, theability for resilience is lost, and the natural disaster be-comes a systemic event.
DISCUSSION
We present a novel approach for estimating the indirecteconomic losses caused by natural disasters by combininga probabilistic physical damage catastrophe model witha macroeconomic ABM. The method has been appliedto flood scenarios in Austria. The ABM is calibratedto the Austrian economy (households, non-financial andfinancial firms and a government sector) at a scale of1:1, i.e. every economic agent in Austria (about 10 mil-lion) is represented in the model. The ABM incorpo-rates an input-output model with 64 industries, whereall goods and services are produced endogenously, anddepicts the relevant constraints regarding productive ca-pacities and financing conditions at the level of individualagents. This allows us to estimate the indirect economiceffects of natural disasters that are caused by the un-folding sequence of economic events following the initialdestruction of productive capital and dwellings. In par-ticular, the model shows the indirect economic effects ofsimulated disaster shocks of controlled size on the Aus-trian economy in terms of cumulative GDP growth andother major macroeconomic variablesThe model produces realistic results in a number ofaspects. Specifically, results correspond well with recentempirical findings of studies including sectoral detail anddifferent types of disasters such as [7, 8, 10, 14, 18, 35].We find that moderate losses due to 100- and 250-yearflood events have small but positive short- to medium-term impacts, while they lead to negative long-term ef-fects of similar magnitude. These results correspond par-ticularly well with [10], which shows that floods of mod-erate magnitude in developed economies, while resultingin positive short- to medium-term effects, have slightlynegative cumulative effects in the long run. Short- tomedium-term results are further supported by a firm-level empirical study on flooding disasters in Europe [14],which finds higher average firm asset and employmentgrowth for regions affected by flooding disasters. Com-parable impacts are also obtained by [7, 8], which reportaggregate short- to medium-term positive growth effectsof floods, and by [18], which shows that natural disas-ters have positive short-term growth effects in developedeconomies. Negative long-term effects for climatic disas-ters are also reported in [35].Our study is the first to estimate the indirect economicimpacts of severe floods in a developed economy on anempirical basis, at the same time taking account of com-plex economic interactions and dynamics with our mod-eling approach. Simulations of severe disasters, such as a1,500-year flood event in Austria, induce a pronouncedlymore negative long-term economic impact on the Aus-trian economy. These results, in line with a theoreti-cal analysis conducted by [42], demonstrate that nega-tive indirect effects from severe disasters are primarilydue to constraints on the productive capacity of the ex-isting capital stock, credit provision and government fi-nances, which impede immediate reconstruction and thusimpose a ceiling on positive growth effects. Empiricalfindings regarding severe floods are largely inconclusive:[7–10, 20, 21] all report no, insignificant or (once con-trolled for) vanishing effects for systemic floods in par-ticular. Especially regarding the most detailed study onflooding disasters [10], a lack of data on severe floods indeveloped countries seemed not to have permitted theobtaining of results for this country group.A unique feature of this analysis is that disaster im-pacts are disaggregated across 64 industry sectors and si- multaneously tracked over time, demonstrating how pos-itive aggregate economic consequences may result in win-ners and losers subject to particular dynamics. The sec-toral decomposition of results reveals that, while somesectors providing the means for the reconstruction of cap-ital stock might profit (predominantly the constructionsector, to a lesser extent the manufacturing sector), oth-ers that are particularly hit by the disaster suffer fromlarge losses that take several years to be compensated(especially the real estate sector). We compute the dis-tribution of losses and their dynamics over time acrosssectors at a higher level of detail than previous stud-ies. Models featuring endogenous dynamics such as CGEmodels, even though they depict up to 35 industry sectorsas in [30], typically are comparative static CGE models [30–32], while fully dynamic CGE models usually depictonly one output good as in [29]. Furthermore, they oftenare confined to regions smaller than a national economy[31, 32]. Sectoral empirical studies such as [7, 8, 10] – be-sides being limited methodologically as set forth above –typically divide the economy into two to three aggregatesectors (agriculture, industry and/or service sector). IOmodels applied to comparable contexts, such as [23], usu-ally exploit the full range of national IO tables (mostlyaround 60-70 industries), but lack the endogenous non-linear dynamics present in the ABM.We show that disasters trigger cyclical economic re-sponses that follow a classic multiplier-accelerator mech-anism as described in [43]. The cycle is caused by an over-shooting of investment during the reconstruction phaseleading to an economic boom, which is followed by adownturn due to a lack of demand once the restora-tion of capital stock has been completed. As reportedin [7, 8, 10], indirect economic effects immediately af-ter the event are predominantly negative due to initiallosses of capital stock, income, as well as demand andsupply of goods. The positive economic stimulus due toreconstruction occurs with a delay of at least a quar-ter, because some time is necessary to compensate forlost capital stock and income. This impulse in turn trig-gers an economic cycle. In the long term, the conse-quences of this cycle tend to outweigh the positive eco-nomic impact induced by reconstruction activities. Sucha cyclical mechanism – which is different from Schum-peter’s creative-destruction or productivity effect, whichhas been the focus of several studies – has received little I.e. they compare an initial equilibrium state before the disasterand another equilibrium state after the disaster without consid-eration of the dynamics between these two economic equilibria. We do not consider this productivity effect in the present study,since its empirical relevance is unclear and subject to extensivedebate in the literature, where different empirical studies presentmixed evidence on growth effects and associated increases in cap-ital productivity after a natural disaster. For further discussionon the productivity effect see [3], for empirical studies presentingpositive growth effects attributed to the productivity effect see[44] and [45], whose findings are contradicted by several otherempirical studies, see [19, 20, 35, 46, 47].
BondsSocial benefitsAdvancesReserves TaxesTaxes LoansDepositsDeposits Taxes ImportsImportsExportsConsumptionDividends Wages,Dividends Subsidies,Consumption - FIG. 5. Schematic overview of the ABM structure showingthe institutional sectors (households, non-financial and finan-cial firms and a general government), and their interactions.The stacked bars show an example of the distributions of di-rect (left) and indirect (right) total losses to the government(white), firms (red), and households (blue). attention in the literature up to now, with the exceptionof a theoretical analysis in [48].This study is the first to combine a probabilistic phys-ical damage catastrophe model with a macroeconomicABM to quantitatively relate disaster sizes with the indi-rect economic impacts. We find a non-trivial behavior ofthe cumulative GDP growth effects as a function of thedirect damage size. We determine a threshold beyondwhich the full productive capacity of an economy hasbeen exploited to restore destroyed capital stock. At thispoint, which is at around 5 % of destroyed capital stock,resilience is lost, and economic growth is dominated bydirect losses. Previous studies such as [42] have hith-erto investigated this matter on a theoretical and moreaggregate basis only.We believe that in times of increased frequency andseverity of potential climate-change-related natural disas-ters, it is expedient to anticipate their short- to long-termeconomic implications, direct as well as indirect ones. Inparticular, it is important to identify potential economiclosers of these events, so as to optimally prepare for afair and efficient post-crisis management.
MATERIALS AND METHODSAgent based model of a small economy
We employ an ABM – developed in [40] and [49] –which depicts the economy of a small nation at a 1:1scale (about 10 million agents) (see SI for a model de-scription). The model is based on detailed data sourcesfrom national accounts, input-output tables, governmentstatistics, census data and business surveys, and is able to closely approximate time series of major macroeco-nomic variables (GDP, inflation, household consumption,investment). The basic structure of the model is depictedin Fig. 5. The model is calibrated to the economy ofAustria in the year 2013, for which the required data isavailable (see SI). Large economies are still out of scopefor such simulations within reasonable computing time.Simulations were carried out on the supercomputer of theVienna Scientific Cluster.
Flood risk estimation and damage scenario generator
We estimate the disaster risk distributions for floodlosses in Austria using a copula approach, and builda damage-scenario generator based on spatially explicitdata to simulate losses to individual households, non-financial and financial firms and government entitiesacross the 64 economic sectors represented in the ABM.The damage-scenario generator simulates a shock to indi-vidual agents in the ABM, which subsequently alter theirbehavior and create higher-order indirect effects over agiven time period (see SI).
ACKNOWLEDGMENTS
We acknowledge support from the EC H2020 projectSmartResilience under grant agreement No 700621.Computations were performed in part on the Vienna Sci-entific Cluster. We would like to thank Prof. Georg Pflugas well as Dr. Anna Timonina-Farkas for helpful com-ments in designing the modeling approach and providinguseful inputs.
REFERENCES [1] J. M. Melillo, T. Richmond, and G. Yohe, Third NationalClimate Assessment (2014).[2] National Research Council,
The impacts of natural dis-asters: A framework for loss estimation (NationalAcademies Press, 1999).[3] S. Hallegatte and V. Przyluski,
The economics of naturaldisasters: concepts and methods , Tech. Rep. 5507 (TheWorld Bank Policy Research Working Paper, 2010).[4] I. Linkov, T. Bridges, F. Creutzig, J. Decker, C. Fox-Lent, W. Kr¨oger, J. H. Lambert, A. Levermann, B. Mon-treuil, J. Nathwani, et al. , Nature Climate Change , 407(2014).[5] H. C. Cochrane, in Modeling Spatial and Economic Im-pacts of Disasters (Springer, 2004) pp. 37–52.[6] C. T. West, in
Proceedings of Analyzing Economic Im-pacts and Recovery from Urban Earthquakes: Issuesfor Policy Makers, Conference presented by EarthquakeEngineering Research Institute and Federal EmergencyManagement Agency, Pasadena, California (1996).[7] N. V. Loayza, E. Olaberria, J. Rigolini, and L. Christi-aensen, World Development , 1317 (2012). [8] T. Fomby, Y. Ikeda, and N. V. Loayza, Journal of Ap-plied Econometrics , 412 (2013).[9] E. Cavallo, S. Galiani, I. Noy, and J. Pantano, Reviewof Economics and Statistics , 1549 (2013).[10] J. Cunado and S. Ferreira, Land economics , 149(2014).[11] E. Cavallo, I. Noy, et al. , International Review of Envi-ronmental and Resource Economics , 63 (2011).[12] C. Kousky, Energy Economics , 576 (2014).[13] N. Shabnam, International Journal of Disaster Risk Sci-ence , 157 (2014).[14] A. M. Leiter, H. Oberhofer, and P. A. Raschky, Envi-ronmental and Resource Economics , 333 (2009).[15] S. Lazzaroni and P. A. van Bergeijk, Ecological Eco-nomics , 333 (2014).[16] J. Klomp and K. Valckx, Global Environmental Change , 183 (2014).[17] P. S. Mohan, B. Ouattara, and E. Strobl, EcologicalEconomics , 1 (2018).[18] I. Noy, Journal of Development economics , 221 (2009).[19] S. Hochrainer, World Bank Policy Research Working Pa-per (2009).[20] I. Noy and A. Nualsri, University of Harvard WorkingPaper (2007).[21] C. R. Jaramillo, Documentos CEDE Universidad de LosAndes (2009).[22] Y. Okuyama and J. R. Santos, Economic Systems Re-search , 1 (2014).[23] S. Hallegatte, Risk analysis , 779 (2008).[24] Y. Okuyama, Economic Systems Research , 115(2007).[25] A. Rose and J. Benavides, Advances in social scienceanalysis of earthquakes, Buffalo, NY: National Centerfor Earthquake Engineering Research (1997).[26] R. V. Whitman, T. Anagnos, C. A. Kircher, H. J. Lago-rio, R. S. Lawson, and P. Schneider, Earthquake Spectra , 643 (1997).[27] R. Boisvert, Indirect economic consequences of a catas-trophic earthquake , 207 (1992).[28] H. Cochrane, Menlo Park, CA: Risk Management Solu-tions (1997).[29] H. Strulik, T. Trimborn, et al. , Natural disasters andmacroeconomic performance , Tech. Rep. (Vienna Uni-versity of Technology, Institute for Mathematical Meth-ods in Economics, Research Group Economics (ECON), 2016).[30] P. K. Narayan, Applied Economics Letters , 721(2003).[31] A. Rose, G. Oladosu, and S.-Y. Liao, Risk Analysis ,513 (2007).[32] A. Rose and S.-Y. Liao, Journal of Regional Science ,75 (2005).[33] D. Brookshire and M. McKee, Indirect Economic Conse-quences of a Catastrophic Earthquake , 267 (1992).[34] R. Boisvert, Report to the Federal Emergency Manage-ment Agency, Ithaca, NY: Cornell University (1995).[35] C. Raddatz, The Wrath of God: Macroeconomic Costs ofNatural Disasters (World Bank Washington DC, 2009).[36] A. Rose, in
Modeling Spatial and Economic Impacts ofDisasters. Advances in Spatial Science , edited by C. S. E.Okuyama Y. (Springer, Berlin, Heidelberg, 2004).[37] N. Bolton and L. Kimbell, in annual meeting of the Pop-ulation Association of America (1995).[38] H. B. Cochrane, B. Kling, D. Black, J. Schmehl,J. Curtsinger, and J. Steenson, in
The Economics ofNatural Hazards (Univeristy of Colorado, Boulder, 1996).[39] C. T. West and D. G. Lenze, International regional sci-ence review , 121 (1994).[40] S. Poledna, M. Miess, and S. Thurner, “Economic fore-casting with an agent-based model,” (2017), in prepara-tion.[41] P. A. Samuelson, Review of Economic Statistics , 75(1939).[42] S. Hallegatte, J.-C. Hourcade, and P. Dumas, Ecologicaleconomics , 330 (2007).[43] P. A. Samuelson, The Review of Economics and Statistics , 75 (1939).[44] J. M. Albala-Bertrand, The Political Economy of LargeNatural Disasters with Special Reference to DevelopingCountries (Oxford: Clarendon Press, 1993).[45] M. Skidmore and H. Toya, Economic Inquiry
Volume40, Issue 4 , 664 (2002).[46] I. Noy, Journal of Development Economics , 221(2009).[47] C. R. H. Jaramillo, Universidad de los Andes, mimeo(2009).[48] S. Hallegatte and M. Ghil, Ecological Economics , 582(2008).[49] T. Assenza, D. Delli Gatti, and J. Grazzini, Journal ofEconomic Dynamics and Control50