Abstract
Dynamic risk processes, which involve interactions at the hazard and risk levels, have yet to be clearly understood and properly integrated into probabilistic risk assessment. While much attention has been given to this aspect lately, most studies remain limited to a small number of sitespecific multirisk scenarios. We present a generic probabilistic framework based on the sequential Monte Carlo Method to implement coinciding events and triggered chains of events (using a variant of a Markov chain), as well as timevariant vulnerability and exposure. We consider generic perils based on analogies with real ones, natural and manmade. Each simulated time series corresponds to one risk scenario, and the analysis of multiple time series allows for the probabilistic assessment of losses and for the recognition of more or less probable risk paths, including extremes or lowprobability–highconsequences chains of events. We find that extreme events can be captured by adding more knowledge on potential interaction processes using in a brickbybrick approach. We introduce the concept of risk migration matrix to evaluate how multirisk participates to the emergence of extremes, and we show that risk migration (i.e., clustering of losses) and risk amplification (i.e., loss amplification at higher losses) are the two main causes for their occurrence.
Introduction
Multirisk assessment is still in its infancy and often only refers to the analysis of multiple single hazards in a same framework (Grünthal et al. 2006; Carpignano et al. 2009; Schmidt et al. 2011). Major catastrophes however remind us that multirisk is not simply the sum of individual risks but that correlations between natural hazards, technological hazards and our complex socioeconomic networks lead to greater risks (e.g., 2005 hurricane Katrina, USA; 2010 eruption of Eyjafjallajökull, Iceland; 2011 Tohoku earthquake, Japan). Innovative methods have been proposed in recent years to tackle the problem of hazard interactions (e.g., Marzocchi et al. 2012) and of other dynamic aspects of risk, such as timedependent vulnerability and exposure (e.g., Selva 2013) or network failures (e.g., Adachi and Ellingwood 2008). However, so far, only a limited number of scenariobased and/or sitespecific multirisk studies have been proposed due to the difficulty and novelty of the task.
Development of a comprehensive multirisk framework is hampered by the following requirements: (1) large amount of input data (2) crossdisciplinary expertise and (3) innovative risk assessment methods. The first two points are generally solved in dedicated multirisk projects at the national (e.g., HAZUSMH, http://www.fema.gov/hazus), international (e.g., CAPRA, http://www.ecapra.org/) or private sector levels (e.g., Grossi and Kunreuther 2005; Schmidt et al. 2011). The third point remains to be solved. As indicated by Kappes et al. (2012), “despite growing awareness of relations between hazards, still neither a uniform conceptual approach nor a generally used terminology is applied”. Similarly, but based on feedback from civil protection stakeholders, Komendantova et al. (2014) noted that: “two areas are most problematic. These are (1) the absence of clear definitions and (2) the lack of information on the added value of multirisk assessment”.
In the present study, we present a novel, generic, multirisk framework based on the sequential Monte Carlo Method (MCM) to allow for a straightforward and flexible implementation of hazard interactions, which may occur in a complex system. Realworld examples of hazard interactions include: earthquake clustering, storm clustering, tsunamis following earthquakes or landslides, landslides or fire following earthquakes, storm surges associated to hurricanes, technological accidents triggered by natural events (i.e., NaTech events). Timevariant vulnerability and exposure related to hazard clustering are also considered, although not the primary focus of this study. More generally, timevariant vulnerability may refer to different processes, such as structure ageing, notrepaired predamage due to past events or damage conditioned on the cooccurrence of several events. Timevariant exposure supposes the evolution of assets value with time, which may be due to socioeconomic factors or to previous losses.
Our goal is specifically to capture and quantify extreme (i.e., lowprobability–highconsequences) events using inductive generalization (e.g., Bier et al. 1999) and by following the recommendation of Kameda (2012), which is to “mobilize “scientific imagination” in the process of decision”—by incorporating extreme events in risk modelling if no observations but sound scientific bases are available (Note that the term “reasoned imagination” is used by PatéCornell 2012). Our approach differs from sitespecific and scenariobased studies (e.g., Adachi and Ellingwood 2008; Marzocchi et al. 2012; Selva 2013) in that we do not define any specific hazard or risk interaction but a framework to implement any type of interaction. With such an objective, real interaction processes have to be abstracted to more basic concepts and engineering methods bypassed. The proposed framework is described in Sect. 2, in which the concept of hazard correlation matrix is introduced.
Validation of our framework, which should be considered as a proofofconcept, is made using generic data and processes defined heuristically. This strategy, that is the use of intuitive judgment and simple rules, allows for the solving of problems that are otherwise difficult to consider. Based on an extensive literature survey, we generalize the concepts of peril, of peril characterization and of hazard interaction. Our aim, by abstracting these concepts into basic categories, is to provide some general guidelines for extreme event quantification. The proofofconcept is presented in Sect. 3, and a discussion on the applicability of the proposed framework to realworld conditions is given in Sect. 4.
Basic hazard and risk terms as well as definitions used in the present article are based, when available, on the book “Catastrophe Modelling: A new approach to managing risk” by Grossi and Kunreuther (2005). Symbols used in our study are listed in Table 1.
Generic multirisk framework
Sequential Monte Carlo Method
The proposed multirisk framework is formed of a core simulation algorithm based on the MCM. We adopt the MCM for its flexibility when dealing with complex systems. We generate N _{sim} time series, sampling events from a Poisson distribution (homogeneous or nonhomogeneous process). Each time series represents one risk scenario, and the analysis of N scenarios allows for the probabilistic assessment of losses and for the recognition of more or less probable risk paths. These risk paths emerge naturally from the system implemented in the MCM.
A set of stochastic events is defined as input for the MCM, with each event characterized by an identifier, a longterm occurrence rate λ and a loss Λ. The loss is defined as Λ = δ Ε with δ the mean damage ratio and Ε the system exposure. δ is derived from the hazard intensity ι with each event being represented by one unique hazard footprint. Each event is therefore implicitly related to one specific source, e.g., an earthquake related to a given fault segment, a storm related to a given track (see Sect. 3.1.1 for the definition of a stochastic event set).
Hazard interactions as well as timevariant exposure and vulnerability represent nonstationary processes, which require additional inputs (see Sects. 3.1.2, 3.1.3) and a sequential processing strategy. The proposed sequential MCM is defined as follows:

Multihazard assessment: define the simulation set with simulation identifier, event identifier and event occurrence time t.

1.
Generate N _{sim} random time series: Sample N _{sim} sets of events over the time interval Δt = [t _{0}; t _{max}] drawn from the Poisson distribution with each stochastic event i characterized by the longterm rate parameters λ _{ i }. Affix an occurrence time t to each event following the random uniform distribution. Record the time series in the simulation set S _{0}, which represents the null hypothesis H _{0} of having no interaction in the system. Fix increment j = 1, which indicates the occurrence of the first event in the time series.

2.
For each of the N _{sim} simulations, record the characteristics of the jth event, which occurs at t _{ j }, in simulation set S _{1}. Resample events k occurring in the interval [t _{ j }; t _{max}] if the conditional probability Pr(kj) exists. This conditional probability is defined in the hazard correlation matrix, described in Sect. 2.2. Affix t _{ k } = t _{ j } + ε with ε ≪ Δt. Fix j = j + 1.

3.
Repeat step 2 while t _{ j } ≤ t _{max}.

4.
Fix j = 1.

1.

Multirisk assessment: update the simulation sets S _{0} and S _{1} with event loss Λ.

5.
For each of the N _{sim} simulations, calculate the mean damage ratio δ _{ j } due to the jth event, which is potentially conditional on the occurrence of previous events. The implementation of timevariant vulnerability is described in Sect. 2.3.

6.
For each of the N _{sim} simulations, calculate the loss Λ _{ j } due to the jth event, which is potentially conditional on the occurrence of previous events. The implementation of timevariant exposure is also described in Sect. 2.3. Record Λ _{ j }.

7.
Repeat steps 5 and 6 while t _{ j } ≤ t _{max}.

5.
Figure 1a illustrates the difference between simulation sets S _{0} and S _{1}. All simulated time series in S _{1} start with the same event as in S _{0}, since changes are only conditional on the occurrence of previous events. Following an event, changes may or may not occur in S _{1} depending on the conditional probabilities and the effects of sampling. Once an event is triggered, the trigger/target pair clusters in time with chains of n events potentially emerging in the time interval [t, t + nε]. The impact of timevariant exposure and vulnerability on event losses are not represented in Fig. 1a. It should be noted that the proposed framework for hazard interactions is a variant of a Markov chain. In that view, the hazard correlation matrix corresponds to a transition matrix, the conditional probabilities to transition probabilities and events to states.
Hazard correlation matrix
We introduce the concept of hazard correlation matrix to quantify hazard interactions. It should be noted that we use a loose definition of the word interaction, as we also refer to oneway causal effects by this term (noted→). The hazard correlation matrix is illustrated in Fig. 1b where trigger events are represented in rows i and target/triggered events in columns j. A given peril P consists of n events P _{ i } with 1 ≤ i ≤ n. Each cell of the square matrix indicates the 1to1 conditional probability of occurrence Pr(ji) = Pr(P _{ j }P _{ i }) over Δt, which is used as input in the MCM. We also consider the nto1 conditional probability by incorporating a memory element to the correlation matrix. Various interaction processes may be implemented, based on empirical, statistical or physical laws. Those are run in the background with only the conditional probability represented in the hazard correlation matrix. However, the memory element is defined such that it can alter the process in the background by informing it of the sequence of previous events. This is illustrated by several examples in Sect. 3.1.2. The approach is different from a strict Markov chain in the fact that it is not memoryless and because the matrix does not require \(\mathop \sum \limits_{j} \Pr \left( {ji} \right) = 1\) (i.e., finite chains of events).
We define various terms (noted in italics) to categorize different types of interactions based on the concept of hazard correlation matrix: An event repeat is described by P _{ i } → P _{ i }, an intrahazard interaction by P _{ i } → P _{ j } and an interhazard interaction by A _{ i } → B _{ j } with A and B two different perils. Moreover, perils can be separated into primary perils A when λ _{ A } > 0 and secondary perils B when λ _{ B } = 0 and Pr(BA) > 0, with λ the longterm occurrence rate. Invisible events i, which do not yield direct losses, should be included in the system if they trigger events j that do (case Pr(ji) > 0, Λ _{ i } = 0 and Λ _{ j } > 0).
Timevariant exposure and timevariant vulnerability
Cumulative losses due to the occurrence of successive events cannot exceed the total exposure Ε. Timevariant exposure can be defined by.
where Ε _{0} the original exposure, Ε _{ i } the exposure of the system immediately following event i and e _{ i } a function representing the exposure reconstruction up to the occurrence time of event i. With a timevariant exposure, the event loss also becomes timevariant with Λ _{ i } = δ _{ i } Ε _{ i−1}. Two endmembers of Eq. 1 are instantaneous reconstruction (i.e., e _{ i } = ∑Λ _{ i }) with Ε _{ i } = Ε _{0} and no reconstruction (i.e., e _{ i } = 0) with 0 ≤ Ε _{ i } ≤ Ε _{0}. Both cases are tested in Sect. 3. Although not considered in the present study, a timedependent term e(t) could be added to Eq. 1 to represent system recovery after a disaster and socioeconomic evolution (e.g., increase in wealth/assets with time).
The clustering of events in time may also influence the vulnerability, which can be described by the conditional mean damage ratio δ _{ ji }. The dependence on the trigger event i may take different forms, independently of the framework developed here. An example of vulnerability dependence on hazard intensity is given in Sect. 3.1.3. Timedependent vulnerability δ _{ i }(t), such as ageing, is not considered.
Proofofconcept
Generic data and processes
We generate generic data and processes by following the heuristic method and by abstracting the concepts of peril, of peril characterization and of hazard interaction into basic categories. Our approach provides some general guidelines for extreme event quantification and a dataset for testing the generic multirisk framework described in Sect. 2. It follows the existing recommendations on extreme event assessment (Bier et al. 1999; Kameda 2012; PatéCornell 2012) by combining inductive generalization and “scientific imagination” to include known examples of extremes as well as potential “surprise” events in a same framework. We intentionally do not consider the case of networks (Adachi and Ellingwood 2008). Data and processes are then implemented in the MCM, and the results analysed in Sects. 3.2 and 3.3. In the present study, we use N _{sim} = 10^{5}, Δt = [t _{0} = 0; t _{max} = 1] and ε = 0.01. It should be noted that the numerous assumptions made below are not a requirement of the proposed multirisk framework but are working hypotheses for basic testing purposes.
Building a stochastic event set
We first generate a stochastic event set in which each event is defined by an identifier, a longterm occurrence rate λ and a loss Λ. Let’s consider peril A consisting of n _{ A } events A _{ i } with 1 ≤ i ≤ n _{ A }. First, we define the frequencyintensity distribution
where ι _{ i } is the hazard intensity of event A _{ i }, λ _{ i } the longterm occurrence rate of event A _{ i } and β the exponential law exponent for peril A. Equation 2 applies to numerous perils, such as earthquakes (Gutenberg and Richter 1944), volcanic eruptions (Simkin and Siebert 1994) or asteroid impacts (Brown et al. 2002). Second, we estimate the mean damage ratio δ of repair cost to the replacement cost of the system from a vulnerability curve that relates the susceptibility of the system to the hazard intensity, with 0 (no damage) ≤ δ ≤ 1 (total destruction). We use the cumulative lognormal distribution
where ι _{ i } is the hazard intensity and δ _{ i } the mean damage ratio due to event A _{ i }. μ and σ are, respectively, the mean and standard deviation of the variable’s logarithm and depend on the peril and asset response type (see Sect. 3.1.3 for a variable μ). Equation 3 is well suited to describe the typical Sshaped vulnerability curve used in damage assessment for earthquakes (Fabbrocino et al. 2005), volcanic eruptions (Spence et al. 2005; Zuccaro et al. 2008), winds (Wehner et al. 2010), fluvial floods (Reese and Ramsay 2010), tsunamis (Srivihok et al. 2012) or asteroid impacts (Mignan et al. 2011). In the present case, we consider a singledamage state, which is an oversimplification compared to standard risk assessment. This limitation is discussed in Sect. 4.
Let’s now consider additional perils other than A. Direct comparison of hazards (socalled joint visualization of multiple hazards) would require the use of a rather subjective hazard intensity classification scheme (e.g., low–medium–high) (Kappes et al. 2012) since intensity ι differs from one peril to another one (e.g., ground shaking, ash load, inundation depth, wind speed, bolide energy). Thus, we remove ι by combining Eqs. 2 and 3, which yields
Using risk as a common language, we resolve the problem of hazard comparability. Event loss Λ _{ i } is then defined by Λ _{ i } = δ _{ i } Ε where Ε = 1 the system exposure. Since a unique vulnerability curve is used for the full exposure, all losses given in the present study are mean loss values.
We generate the stochastic event set based on Eq. 4 with 10^{−4} ≤ λ _{ i } ≤ 10^{−1} in 0.1 increments in the log scale. It follows that 31 events are defined per peril with return periods varying from 10 to 10,000 years. This applies only to primary perils (here A and B), which occur spontaneously following Eq. 2. We fix β _{ A } = ln(10) and β _{ B } = 0.5ln(10) (Fig. 2a). Perils A and B are analogue to earthquakes and volcanic eruptions for instance. Secondary perils (here C, D and E), by definition, have a null longterm occurrence rate λ = 0 and only occur following primary events. Examples include tsunamis following earthquakes (Suppasri et al. 2012), storm surges (Irish et al. 2008) or Natech events (i.e., industrial accidents with natural hazard triggers, Krausmann et al. 2011). Vulnerability curve parameters (Eq. 3, Fig. 2b) are fixed such that perils A and B show distinctive risks (i.e., different risk ranking), with peril A dominating risk at short return periods and peril B dominating risk at long return periods (Eq. 4, Fig. 2c). We use μ _{ A } = ln(5), σ _{ A } = 0.4, μ _{ B } = ln(6) and σ _{ B } = 0.1. We define three events per secondary peril (C _{1}–C _{3}, D _{1}–D _{3} and E _{1}–E _{3}) and use a same vulnerability step function such that δ _{1} = 0.01, δ _{2} = 0.1 and δ _{3} = 1 for the 3 perils. Table 2 shows the resulting stochastic event set, which is used as input in the MCM.
Populating the hazard correlation matrix
Providing a comprehensive review of hazard interactions is out of the scope of this study. Such a daunting task was for example described by Gill and Malamud (2012) (see also Kappes et al. 2012). Here, we describe the principal types of hazard interactions that can be implemented in the correlation matrix to be used for proofofconcept of the proposed multirisk framework. The proposed hazard correlation matrix is shown in Fig. 3. We consider the three categories of interactions previously defined: event repeat (e.g., A _{ i } → A _{ i }; C → C), intrahazard interaction (e.g., A _{ i } → A _{ j }) and interhazard interaction (e.g., A _{ i } → B _{ j }). The effect can be positive (i.e., probability increase) or negative (i.e., probability decrease),, and temporary or lasting (concept of memory). Table 3 lists the different combinations defined in Fig. 3 and examples of analogies with real perils. More types of interactions could be envisioned, as not all possible combinations are described here (primary vs. secondary/interaction category/effect/memory or not).
Event repeat for primary perils is here described by the lognormal distribution
where λ _{ i } is the longterm occurrence rate of event A _{ i } and α _{ i } = 1 the shape. In Eq. 5, the probability over Δt is very low immediately following t _{0} for Δt ≪ 1/λ _{ i } and progressively increases through time. It hence represents a renewal process in which energy is first released by an event and then accumulates gradually through time due to a loading process. It applies for example to earthquakes, which occurrence is controlled by tectonic loading (Parsons 2005). We apply Eq. 5 to perils A and B, which significantly lowers the probability of occurrence compared to the longterm probability Pr(A _{ i }) = 1 − exp(−λ _{ i } Δt) for Δt ≪ 1/λ _{ i } (Fig. 4). Event repeat for secondary perils is described by Pr(C_{ j }C _{ i }) = Pr(D_{ j }D _{ i }) = Pr(E_{ j }E _{ i }) = 0, which indicates that a given peril can only occur once, whether immediately after the event (case of peril C—no memory) or through Δt (case of perils D and E—with memory) (Fig. 3; Table 3). Peril C is analogue to a tsunami or storm surge, which can occur anytime the proper trigger occurs, but only once per triggering. Perils D and E are analogue to technological accidents. If such an accident yields nonfunctionality, the same accident cannot occur again if the critical infrastructure is not repaired. If it were repaired immediately, it would then be in the peril C class (Table 3). Memory is here defined by the recording of previous instances of event occurrences.
We consider that primary perils are due to an underlying loading process, as described previously for repeating events by Eq. 5. In this case, events are advanced or delayed from a time shift ΔT _{ ij }. We define
based on the nonhomogenous Poisson process with i the trigger event, j the target event and
ΔT _{ ij } > 0 represents a time delay. λ _{ j,mem} is updated after each event occurrence j, including the case i = j, and represents the memory of the process (at t _{0}, λ _{ j,mem} = λ _{ j }). The same approach is used in earthquake interaction modelling, which is well established with the theory of stress transfer (Toda et al. 1998; King 2007). This theory also applies to earthquake/volcanic eruption interactions (Hill et al. 2002; Eggert and Walter 2009). Here, we apply Eqs. 6 and 7 to intrahazard interactions (A _{ i } → A _{ j }; B _{ i } → B _{ j }) and to interhazard interactions (A _{ i } → B _{ j }; B _{ i } → A _{ j }) (Fig. 2). We define the matrix ΔT _{ ij } = ± f/λ _{ i } where f an ad hoc coupling factor, such that the time shift due to event i is proportional to the return period of this event (i.e., a rare large event will have a greater effect than a small, more common, event). Since the spatial relationship between different events is not considered, we simply fix the sign of ΔT _{ ij } randomly for each target event j. We fix f = 0.1 (the role of different values is discussed in Sect. 3.3).
Secondary perils occur based on the following oneway causal effects: A → C, C → D and D → E (Fig. 3). Hence, we can produce the domino effect A → C → D → E. We consider a linear relationship between trigger event i and target event j, such that
where n _{i} and n _{ j } are the number of stochastic events i and j, respectively. This is described in Fig. 5a for the cases A → C and C → D (same for D → E). It shows that for any event i, in the range [1, n _{ i }], there is an associated event j in the range [1, n _{ j }]. Using a relationship between event increments is however artificial. For real perils, the relationship would normally relate hazard intensities ι _{ i } to ι _{ j }. Examples include the relationship h = 3.10^{−5} M ^{6.2344} between tsunami height h and earthquake magnitude M (Suppasri et al. 2012) or the relationship between storm surge height and maximum wind speed, as defined in the Saffir–Simpson hurricane scale (e.g., Irish et al. 2008). Such simple relationships are controversial, since effects depend on site conditions. The concepts remain however valid at a more abstract level. In the generic case presented here, we relate event increments directly for sake of simplicity. To determine the conditional probability Pr(ji), we assume a binomial distribution
where 0 ≤ k ≤ n _{ j }. k = j except for the case k = 0, which corresponds to the probability of having no event triggered Pr(Øi). Pr(ki) is shown in Fig. 5b. Note that the higher the trigger increment i is (i.e., proxy to hazard intensity), the greater is the probability of triggering a severe target event j; which is a direct consequence of Eq. 8.
Considering timevariant vulnerability
Hazard clustering may also influence the vulnerability of the system. One example, which has recently been considered, is the increase in vulnerability to ground shaking due to increased structural load following an ash fall (Zuccaro et al. 2008; Selva 2013). Another wellknown case is the increased vulnerability of structures to successive shakings during an earthquake cluster (Jalayer et al. 2010). Here, we consider the conditional vulnerability curve
where δ _{ ji } the mean damage ratio due to event j conditional on the occurrence of event i and μ _{ ji } = μ _{ j } − γln(ι _{ i }). ι _{ i } and ι _{ j } are the hazard intensities of events i and j, respectively, and γ is an ad hoc calibration parameter. The proposed generic relationship is based on the idea that the higher the intensity ι _{ i } is, the greater is the vulnerability to event j. The validity of such formulation remains to be verified with the logarithmic term here only proposed for consistency within Eq. 10. For real data, a different relationship may be obtained empirically or from structural engineering (Zuccaro et al. 2008; Jalayer et al. 2010). Here, we apply Eq. 10 to perils A and B for the case δ _{ AB } where γ = 0.3. Figure 6 shows the vulnerability curve linked to peril A conditional on the occurrence of events B _{1} to B _{31}. This process is analogue to increased vulnerability to earthquake shaking due to the occurrence of a volcanic eruption with ash fall (Zuccaro et al. 2008; Selva 2013).
Definition of extremes
Extreme events may be defined as events, which are “rare, severe and outside the normal range of experience of the system in question” (Bier et al. 1999). This definition however assumes that extreme events are somewhat anomalous. In this line of reasoning, quantification of extremes would make them normal, i.e., nonextreme. In the present study, extreme events are simply defined as lowprobability–highconsequences events, whether they seem normal or abnormal. Here, extremes do not only refer to individual events, but also to groups of events of which only the overall impact is considered. Therefore, the definition of an event remains blurry and depends on the level considered in the system (i.e., an event may be composed of subevents and a metaevent of events). Then, extremes are categorized into individual events (Sect. 3.2.1) and multiple events, i.e., coinciding events or triggered chains of events (Sect. 3.2.2).
In recent years, anomalous events, or outliers, have been repackaged into fancier animals, some of which made their way into popular culture. Taleb (2007) coined the term “black swan” to describe rare events, which in principle cannot be anticipated. Another popular term is “perfect storm”, which refers to an event resulting of a rare combination of circumstances (PatéCornell, 2012). Finally, the concept of “dragon king” introduced by Sornette and Ouillon (2012) explains that outliers are due to a physical mechanism not represented in the distribution tail considered. These different terms are not used in the present study.
Individual extreme events: concept of heavy tail
Let’s first consider the role of individual events on the overall risk. Figure 7 shows the metric λ _{ i } Λ _{ i }. (loss over Δt) for events i characterized by their longterm occurrence rate λ _{ i } and loss Λ _{ i } (Table 2). A related metric is the annual average loss (AAL). As described in Sect. 3.1.1, peril A dominates the risk at higher frequencies, while peril B dominates at lower frequencies. The highest loss over period Δt is due to event A _{15} (at λ = 0.0039) for peril A and to event B _{21} (at λ = 0.0010) for peril B. We see that although we consider most of the risk in our stochastic event set, we miss a nonnegligible part below λ = 10^{−4}. It is common practice in insurance industry to only consider risk scenarios with a return period that does not exceed 1,000 years (Smolka 2006). In our example and with Δt = 1 year, we would miss most of the risk using such an arbitrary choice (grey area in Fig. 7). This illustrates the concept of heavy tail from which extremes are populated. This tail is however bounded whether by the maximum possible event size (e.g., concept of earthquake maximum magnitude in seismic hazard assessment) or by the exposure (i.e., loss saturation).
Rare unknown perils could also contribute to overall risk, showing the potential instability of the risk process, as any added information on possible extreme events could significantly, if not dramatically, alter the risk measure (Fig. 7). The level of knowledge on extreme individual events is directly linked to the length of the available records (Smolka 2006). New approaches, such as the study of myths (Piccardi and Masse 2007) or of odd geomorphological structures (e.g., Scheffers et al. 2012) within the scope of scientific imagination, should help improving these records and reassess the overall risk. This is however out of the scope of the present study.
Coinciding events and triggered chains of events
Implementation and evaluation of coinciding events and triggered chains of events is the main purpose of this work. Coinciding events are independent events, which occur in cluster by chance. Such behaviour is in principle trivial to treat (if all events are known—see previous section), since it emerges naturally from the Poisson process. This is illustrated in Fig. 8 where the distribution of the number of event occurrences k per simulation for simulation set S _{0} (homogeneous Poisson process, null hypothesis H _{0}) is shown in light grey. With the event rate λ _{tot} = ∑λ = 0.97 (Table 2), we find for instance that the probability of having k = 5 events occurring during the same simulation is Pr(k = 5) ~ 0.003. Large sets of coinciding events are however controlled by the moderatesize events with low return periods. More damaging events are rarer and their random cooccurrence even rarer.
Chains of events are the result of interaction processes. This means that their clustering is nonrandom. Here we consider chains of events by adding knowledge to the system by implementing the generic interaction processes described in Sects. 2.2 (hazard contribution) and 2.3 (risk contribution). We test four different hypotheses H _{1} to H _{4} (four simulation sets S _{1} to S _{4}), as defined in Table 4. These whatif scenarios are as follows: primary peril interactions between events of type A and B (H _{1}); primary peril interactions and timevariant vulnerability linked to the clustering of events of type A and B (H _{2}); primary and secondary peril interactions (including A → C → D → E chains) and timevariant vulnerability (H _{3}); and finally, primary and secondary peril interactions, timevariant vulnerability and timevariant exposure (i.e., no reconstruction) (H _{4}). All these hypotheses are based on the data and processes defined in Sect. 3.1. Results are discussed below.
Emergence of extremes
Risk migration
In hypothesis H _{1}, events from perils A and B interact following the rules described previously (Fig. 3; Eqs. 6–7). The number of events per simulation is compared to the one expected in the null hypothesis H _{0} in Fig. 8. Using the akaike information criterion (AIC), we find that the number of event occurrences k per simulation for simulation set S _{1} (in dark grey) is better described by the negative binomial distribution than by the Poisson distribution, which indicates overdispersion (i.e., clustering of events). Here, we use a coupling factor f = 0.1 (see Sect. 3.1.2 for the definition of f), which gives the index of dispersion (variance/mean ratio) ϕ = 4.7. 0 < ϕ < 1 represents underdispersion and ϕ > 1 overdispersion. A higher f yields a higher ϕ, more coupling yielding more clustering. Let’s note that one could directly simulate risk scenarios by sampling from the negative binomial distribution instead of from the Poisson distribution in step 1 of the MCM and by not applying steps 2–3 (see Sect. 2.1). This approach applies when clustering is not due to event interactions but to a higherlevel process. The negative binomial distribution is frequently used in storm modelling for instance (Mailier et al. 2006; Vitolo et al. 2009). This is however not tested in the present study.
In Fig. 8, the probability of having k = 5 events is Pr(k = 5) ~ 0.04 in set S _{1} in contrast with Pr(k = 5) ~ 0.003 in set S _{0}. For k = 7, Pr(k = 7) ~ 0.02 in set S _{1} while Pr(k = 7) ~ 0.0004 in set S _{0}. With N _{sim} = 10^{5}, we find a maximum number of events per simulation k _{max} = 7 in set S _{0} and k _{max} = 41 in set S _{1}. For hypothesis H _{3}, in which the domino effect A → C → D → E is added (Fig. 3), the index of dispersion increases to ϕ = 7.0, and we get Pr(k = 5) ~ 0.05, Pr(k = 7) ~ 0.03 and k _{max} = 49 in set S _{3}. This demonstrates that risk migrates to lowerprobability–higherconsequences events when hazard interactions are considered. Here, we talk about metaevents, the term consequence being defined as the aggregated loss over the k events of the cluster (assuming in a first time a homogeneous distribution of losses Λ over k).
Risk amplification
As defined in Sect. 3.1.2, the time advance or delay ΔT _{ ij } of events j due to a given trigger i increases with the trigger intensity ι (for perils A and B—hypothesis H _{1}). Therefore, the rarer the event is, the stronger is the associated clustering and the higher is the aggregated loss. Similarly, the higher the intensity ι of a trigger is, the higher is the probability of a larger secondary event (for perils C, D and E; Fig. 5—hypothesis H _{3}). Finally, the higher the intensity of peril B is, the higher the loss due to peril A is (Fig. 6—hypothesis H _{2}). This is illustrated in Fig. 9, which shows the mean aggregated losses (Fig. 9a) and the median aggregated losses (Fig. 9b) observed in simulations where the largest event from peril A is A _{ i }. We find that by adding more information on the dynamic risk process, losses tend to be amplified with increasing risk. We can fit the increase in aggregated losses by a powerlaw relationship with exponent 2.9 and 3.6 for mean and median aggregated losses, respectively. In comparison, the powerlaw exponent is 2.7 for individual event losses (with Λ _{ i } ~ 3.10^{−5} i ^{2.7}). This indicates a phenomenon of risk amplification for lowprobability–highconsequences events.
Let’s finally remark that aggregated losses saturate to ∑Λ = Ε _{0} = 1 in hypothesis H _{4} if the exposure is not reconstructed after a loss. Moreover, the renewal process of event repeats limits the number of occurrences of any given event (Figs. 3, 4; hypotheses H _{1} and H _{3}). Such processes counteract risk amplification and avoid the emergence of exploding chain reactions. Since the data and processes defined in Sect. 3.1 were carefully selected to be representative of existing perils and interactions based on an extensive survey of the literature, our results (risk migration and amplification) are believed to represent some common characteristics of multirisk. It is evident that both aspects strongly depend on site conditions and that only the analysis of real sites will permit to determine the real impact of multirisk. We have demonstrated that the proposed framework could be used for such a task.
A multirisk metric: the risk migration matrix
We can evaluate the role of multirisk in the emergence of extremes using two different standard metrics: the exceedance probability (EP) curve (e.g., Grossi et al. 2005; Smolka 2006) and the risk matrix (Cox 1998; Krausmann et al. 2011). Both are different representations of the relationship between event frequency and event loss (in the present study, between frequency of metaevents and aggregated loss). Figure 10 shows EP curves of the five different simulation sets (Table 4). We see the progressive increase in the area below the curve and the tail becoming fatter when more information on the risk process is added. We also see the case of a bounded EP at ∑Λ = Ε _{0} = 1. However, such a metric provides only limited information on the extreme events that populate the tail of the EP curve.
In contrast, the risk matrix provides a more visual representation of the risk although the colour code, from green (minimum risk) to red (maximum risk) (Krausmann et al. 2011), remains subjective (Fig. 11, top left). We populate the risk matrix with the risk scenarios or risk paths generated in the different simulation sets. Each simulation is characterized by an aggregated loss, which represents one risk scenario. The number of times N the same aggregated loss is observed gives the frequency N/N _{sim} of this risk scenario. Here, aggregated losses are binned into 0.01 cells in the log scale (i.e., 501 different values in the loss interval 10^{−4} ≤ Λ ≤ 10). Each risk scenario is then represented by a point in the risk matrix (Fig. 11). Using a smaller bin would shift all values to lower frequencies, but the observed patterns would remain the same. This approach should therefore be considered as semiquantitative.
To better evaluate the role of multirisk in the emergence of extremes, we introduce the notion of risk migration matrix, which is shown in Fig. 11. It is defined as the difference in the density of risk scenarios observed between two hypotheses. To avoid a pixellated result, the densities are first calculated using a Gaussian kernel, here with a standard deviation large enough to focus on the firstorder migration patterns. The right column shows the case H _{ i } − H _{0}, and the left column shows the case H _{ i } − H _{i−1} with i the hypothesis number (Table 4). Risk scenarios of the first and second hypotheses are represented by white and black points, respectively. An increase in risk is represented in red and a decrease in blue. The proposed approach allows us to visualize how the risk migrates as a function of frequency and aggregated losses when new information is added to the system. From H _{0} to H _{3}, we see a progressive shift of the risk to higher frequencies and higher losses (from yellow to red), indicating that the risk is underestimated when interactions at the hazard and risk levels are not considered. The shift is particularly pronounced in the case of domino effects with A → C → D → E. Finally, the risk migration matrix H _{4} − H _{3} shows how timedependent exposure yields a saturation of losses. Feedback from civil protection stakeholders also showed that a risk matrix view might be preferable to the use of loss curves for communicating multirisk results. This is in the context of this feedback that the concept of risk migration matrix was developed (Komendantova et al. 2014).
Applicability to realworld conditions
We have presented a novel, generic and flexible multirisk framework, which implements coinciding events, triggered chains of events, timevariant vulnerability and timevariant exposure. While the sequential MCM is a wellestablished method to model complex systems, it is still rarely used as the basis of a multirisk assessment tool. Figure 12 shows the difference between standard risk modelling (e.g., Grossi et al. 2005) and the newly proposed approach (Sect. 2). Application of the approach to real test sites will require two improvements: (1) a data switch with the transition from generic to real data and processes, which will be the subject of a companion manuscript by the same authors: “The Quantification of LowProbability–HighConsequences Events: Part II. Guidelines to MultiRisk Assessment Based on the Virtual City Concept” and (2) an integration of engineering models that are in use in existing risk tools (e.g., moving from a singledamage state equation (Eq. 3) to more realistic multidamage states as described for instance in Lagomarsino and Giovinazzi 2006). This second requirement is nontrivial, not only due to the different modelling structure, but also to the use here of a stochastic event set instead of aggregated hazard. Aggregated hazard is used for instance in standard seismic risk assessment (e.g., Cornell 1968) and in its recent multirisk extensions (Marzocchi et al. 2012; Selva 2013). Nonetheless, moving to an eventbased sequential simulation approach seems appropriate to model dynamic risk processes and extremes.
Testing of the proposed multirisk framework has been made possible thanks to the heuristic method and the definition of generic perils and processes, which is based on the idea of “scientific imagination” (Kameda 2012; PatéCornell 2012). We have shown that the hazard correlation matrix, by adding knowledge of potential interaction processes, allows for the capture of lowprobability–highconsequences events. In particular, we have shown the role of risk migration and of risk amplification for their occurrence. The present work should be seen as a proofofconcept as we did not attend to fully resolve the difficult problem of extremes. We only considered a selected number of possible interactions, but while the chains of events that emerge in the system may seem obvious, adding more perils and more interactions will yield more complex risk patterns. We thus recommend a brickbybrick approach to the modelling of multirisk, to progressively reduce epistemic uncertainties. A more realistic modelling of lowprobability–highconsequences events will also require the consideration of additional aspects, such as uncertainties (e.g., Barker and Haimes 2009), domino effects in socioeconomic networks (e.g., Adachi and Ellingwood 2008; Buldyrev et al. 2010) and longterm processes (Grossi and Kunreuther 2005; Smolka 2006), such as climate change (Garcin et al. 2008), infrastructure ageing (Rao et al. 2010) and exposure changes (Bilham 2009). While the concepts developed in the present study can suggest the theoretical benefits of multirisk assessment, identifying their realworld practicality will require the application of the proposed framework to real test sites.
References
Adachi T, Ellingwood BR (2008) Serviceability of earthquakedamaged water systems: effects of electrical power availability and power backups systems on system vulnerability. Reliab Eng Syst Saf 93:78–88. doi:10.1016/j.ress.2006.10.014
Barker K, Haimes YY (2009) Assessing uncertainty in extreme events: applications to riskbased decision making in interdependent infrastructure sectors. Reliab Eng Syst Saf 94:819–829. doi:10.1016/j.ress.2008.09.008
Bier VM, Haimes YV, Lambert JH, Matalas NC, Zimmerman R (1999) A survey of approaches for assessing and managing the risk of extremes. Risk Anal 19:83–94
Bilham R (2009) The seismic future of cities. Bull Earthq Eng. doi:10.1007/s1051800991470
Brown P, Spalding RE, ReVelle DO, Tagliaferri E, Worden SP (2002) The flux of small nearEarth objects colliding with the Earth. Nature 420:294–296
Buldyrev SV, Parshani R, Paul G, Stanley HE, Havlin S (2010) Catastrophic cascade of failures in interdependent networks. Nature 464:1025–1028. doi:10.1038/nature08932
Carpignano A, Golia E, Di Mauro C, Bouchon S, Nordvik JR (2009) A methodological approach for the definition of multirisk maps at regional level: first application. J Risk Res 12:513–534. doi:10.1080/13669870903050269
Cornell C (1968) Engineering seismic risk analysis. Bull Seismol Soc Am 59:1583–1606
Cox T (1998) Risk integration and decisionmaking. In: Kirchssteiger C, Christou MD, Papadakis GA (eds) Risk assessment and management in the context of the Seveso II Directive. Industrial Safety Series, 6. Elsevier, Amsterdam
Eggert S, Walter TR (2009) Volcanic activity before and after large tectonic earthquakes: observations and statistical significance. Tectonophysics 471:14–26. doi:10.1016/j.tecto.2008.10.003
Fabbrocino G, Iervolino I, Orlando F, Salzano E (2005) Quantitative risk analysis of oil storage facilities in seismic areas. J Hazardous Mater A123:61–69. doi:10.1016/j.jhazmat.2005.04.015
Garcin M, Desprats JF, Fontaine M, Pedreros R, Attanayake N, Fernando S, Siriwardana CHER, De Silva U, Poisson B (2008) Integrated approach for coastal hazards and risks in Sri Lanka. Nat Hazards Earth Syst Sci 8:577–586
Gill JC, Malamud BD (2012) Reviewing and visualising interaction relationships for natural hazards, EGU General Assembly 2012, 22–27 April, 2012, Vienna, Austria, 708
Grossi P, Kunreuther H (eds) (2005) Catastrophe modeling: a new approach to managing risk. Springer, Berlin
Grossi P, Kunreuther H, Windeler D (2005) Chapter 2—an introduction to catastrophe models and insurance. In: Grossi P, Kunreuther H (eds) Catastrophe modeling: a new approach to managing risk. Springer, Berlin, pp 23–42
Grünthal G, Thieken AH, Scharz J, Radtke KS, Smolka A, Merz B (2006) Comparative risk assessments for the city of cologne—storms, floods, earthquakes. Nat Hazards 38:21–44. doi:10.1007/s1106900585980
Gutenberg B, Richter CF (1944) Frequency of earthquakes in California. Bull Seismol Soc Am 34:185–188
Hill DP, Pollitz F, Newhall C (2002) Earthquake–volcano interactions. Phys Today 55(11):41–47
Irish JL, Resio DT, Ratcliff JJ (2008) The influence of storm size on hurricane surge. J Phys Oceanogr 38:2003–2013. doi:10.1175/2008JPO3727.1
Jalayer F, Asprone D, Prota A, Manfredi G (2010) A decision support system for postearthquake reliability asessment of structures subjected to aftershocks: an application to L'Aquila earthquake, 2009. Bull Earthq Eng. doi:10.1007/s1051801092306
Kameda H (2012), Engineering agenda on nuclear safety from the 2011 TohokuPacific earthquake. In: Proceedings of international symposium engineering lessons learned from the 2011 great East Japan Earthquake, March 1–4, 2012, Tokyo, Japan, 102–121
Kappes MS, Keiler M, von Elverfeldt K, Glade T (2012) Challenges of analyzing multihazard risk: a review. Nat Hazards 64:1925–1958. doi:10.1007/s1106901202942
King GCP (2007) Fault interaction, earthquake stress changes, and the evolution of seismicity. In: Schubert G (ed) Treatise on geophysics, vol 4. Elsevier, Oxford, pp 225–256
Komendantova N, Mrzyglocki R, Mignan A, Khazai B, Wenzel F, Patt A, Fleming K (2014) Multihazard and multirisk decision support tools as a part of participatory risk governance: Feedback from civil protection stakeholders. Int J Disaster Risk Reduct 8:50–67. doi:10.1016/j.ijdrr.2013.12.006
Krausmann E, Cozzani V, Salzano E, Renni E (2011) Industrial accidents triggered by natural hazards: an emerging risk issue. Nat Hazards Earth Syst Sci 11:921–929. doi:10.5194/nhess119212011
Lagomarsino S, Giovinazzi S (2006) Macroseismic and mechanical models for the vulnerability and damage assessment of current buildings. Bull Earthq Eng 4:415–443. doi:10.1007/s105180069024z
Mailier PJ, Stephenson DB, Ferro CAT, Hodges KI (2006) Serial clustering of extratropical cyclones. Mon Weather Rev 134:2224–2240
Marzocchi W, GarciaAristizabal A, Gasparini P, Mastellone ML, Di Ruocco A (2012), Basic principles of multirisk assessment: a case study in Italy. Nat Hazards. doi:10.1007/s110690120092x
Mignan A, Grossi P, MuirWood R (2011) Risk assessment of Tunguskatype airbursts. Nat Hazards 56:869–880. doi:10.1007/s1106901095973
Parsons T (2005), Significance of stress transfer in timedependent earthquake probability calculations. J Geophys Res 110:B05S02. doi:10.1029/2004JB003190
PatéCornell E (2012) On “Black Swans” and “Perfect Storms”: risk analysis and management when statistics are not enough. Risk Anal 32:1823–1833. doi:10.1111/j.15396924.2011.01787.x
Piccardi L, Masse WB (eds) (2007) Myth and geology. Geological Society, London, Special Publications, 273
Rao AS, Lepech MD, Kiremidjian AS, Sun XY (2010) Time varying risk modeling of deteriorating bridge infrastructure for sustainable infrastructure design, 5th international conference on bridge maintenance safety management and life cycle optimization. CRC Press, Taylor & Francis Group, 2494–2501
Reese S, Ramsay D (2010) RiskScape: flood fragility methodology. NIWA Technical Report: WLG201045, p 46
Scheffers A, Scheffers S, Squire P (2012) Trimlines as evidence for palaeotsunamis. Nat Hazards 63:165–179. doi:10.1007/s1106901096916
Schmidt J, Matcham I, Reese S, King A, Bell R, Henderson R, Smart G, Cousins J, Smith W, Heron D (2011) Quantitative multirisk analysis for natural hazards: a framework for multirisk modelling. Nat Hazards 58:1169–1192. doi:10.1007/s110690119721z
Selva J (2013) Longterm multirisk assessment: statistical treatment of interaction among risks. Hazards, Nat. doi:10.1007/s1106901305999
Simkin T, Siebert L (1994) Volcanoes of the world, 2nd edn. Geoscience Press, Tucson
Smolka A (2006) Natural disasters and the challenge of extreme events: risk management from an insurance perspective. Philos Trans R Soc A 364:2147–2165. doi:10.1098/rsta.2006.1818
Sornette D, Ouillon G (2012) Dragonkings: mechanisms, statistical methods and empirical evidence. Eur Phys J Spec Top 205:1–26. doi:10.1140/epjst/e2012015595
Spence RJS, Kelman I, Zuccaro PJ, Petrazzuoli S (2005) Residential building and occupant vulnerability to tephra fall. Nat Hazards Earth Syst Sci 5:477–494. doi:10.5194/nhess54772005
Srivihok P, Honda K, Ruangrassamee A, Muangsin V, Naparat P, Foytong P, Promdumrong N, Aphimaeteethomrong P, Intavee A, Layug JE, Kosin T (2012) Development of an online tool for tsunami inundation simulation and tsunami loss estimation. Cont Shelf Res. doi:10.1016/j.csr.2012.08.021
Suppasri A, Imamura F, Koshimura S (2012) Tsunamigenic ratio of the Pacific Ocean earthquakes and a proposal for a Tsunami Index. Nat Hazards Earth Syst Sci 12:175–185. doi:10.5194/nhess121752012
Taleb NN (2007) The Black Swan: the impact of the highly improbable. Random House, New York
Toda S, Stein RS, Reasenberg PA, Dieterich JH, Yoshida A (1998) Stress transferred by the 1995 Mw = 6.9 Kobe, Japan, shock: effect on aftershocks and future earthquake probabilities. J Geophys Res 103(B10):24543–24565
Vitolo R, Stephenson DB, Cook IM, MitchellWallace K (2009) Serial clustering of intense European storms. Meteorol Zeitschrift 18:411–424
Wehner M, Ginger J, Holmes J, Sandland C, Edwards M (2010) Development of methods for assessing the vulnerability of Australian residential building stock to severe wind, IOP Conf Ser Earth Environ Sci 11, doi:10.1088/17551315/11/1/012017
Zuccaro G, Cacace F, Spence RJS, Baxter PJ (2008) Impact of explosive eruption scenarios at Vesuvius. J Volcanol Geotherm Res 178:416–453. doi:10.1016/j.jvolgeores.2008.01.005
Acknowledgments
We thank two anonymous reviewers for their comments. The research leading to these results has been supported by the New MultiHAzard and MulTiRIsK Assessment MethodS for Europe (MATRIX) project, funded by the European Community’s Seventh Framework Programme [FP7/20072013] under Grant Agreement No. 265138.
Author information
Affiliations
Corresponding author
Rights and permissions
Open Access This article is distributed under the terms of the Creative Commons Attribution License which permits any use, distribution, and reproduction in any medium, provided the original author(s) and the source are credited.
About this article
Cite this article
Mignan, A., Wiemer, S. & Giardini, D. The quantification of lowprobability–highconsequences events: part I. A generic multirisk approach. Nat Hazards 73, 1999–2022 (2014). https://doi.org/10.1007/s1106901411784
Received:
Accepted:
Published:
Issue Date:
Keywords
 Multihazard
 Multirisk
 Extreme event
 Monte Carlo
 Markov chain