# Direct and Hierarchical Models for Aggregating Spatially Dependent Catastrophe Risks

^{*}

## Abstract

**:**

## 1. Introduction

## 2. Copula Trees

#### 2.1. Problem Statement

#### 2.2. Direct Model

#### 2.3. Hierarchical Model

#### 2.4. Implementation of Risk Aggregation at Branching Nodes

#### 2.4.1. Split-Atom Convolution

Algorithm 1: Split-Atom Convolution: 9-products |

Input: Two discrete pdfs ${p}_{X}$ and ${p}_{Y}$ with supports:$x=\{{x}_{1},{x}_{2},{x}_{2}+{h}_{x},{x}_{2}+2{h}_{x},\dots ,{x}_{2}+({N}_{x}-3){h}_{x},{x}_{{N}_{x}}\}$; $y=\{{y}_{1},{y}_{2},{y}_{2}+{h}_{y},{y}_{2}+2{h}_{y},\dots ,{y}_{2}+({N}_{y}-3){h}_{y},{y}_{{N}_{y}}\}$; and probabilities: ${p}_{X}(x)={\sum}_{i=1}^{{N}_{x}}\delta (x-{x}_{i}){p}_{X}({x}_{i})$; ${p}_{Y}(y)={\sum}_{i=1}^{{N}_{y}}\delta (y-{y}_{i}){p}_{Y}({y}_{i})$ ${N}_{s}$ - maximum number of points for discretizing convolution grid - 1:
- ${x}^{(1)}=\{{x}_{1}\}$; ${p}_{X}^{(1)}(x)=\delta (x-{x}^{(1)}){p}_{X}({x}_{1})$
`// split the left atom of ${p}_{X}$` - 2:
- ${x}^{(2)}=\{{x}_{{N}_{x}}\}$; ${p}_{X}^{(2)}(x)=\delta (x-{x}_{i}^{(2)})\xb7{p}_{X}^{(2)}({x}_{i}^{(2)})$
`// split the right atom of ${p}_{X}$` - 3:
- ${x}^{(3)}=\{{x}_{2},{x}_{2}+{h}_{x},{x}_{2}+2{h}_{x},\dots ,{x}_{2}+({N}_{x}-3){h}_{x}\}$; ${p}_{X}^{(3)}(x)={\sum}_{i=1}^{{N}_{x}-2}\delta (x-{x}_{i}^{(3)}){p}_{X}^{(3)}({x}_{i}^{(3)})$
- 4:
- ${y}^{(1)}=\{{y}_{1}\}$; ${p}_{Y}^{(1)}(y)=\delta (y-{y}^{(1)}){p}_{Y}({y}_{1})$
`// split the left atom of ${p}_{Y}$` - 5:
- ${y}^{(2)}=\{{y}_{{N}_{y}}\}$; ${p}_{Y}^{(2)}(y)=\delta (y-{y}_{i}^{(2)})\xb7{p}_{Y}^{(2)}({y}_{i}^{(2)})$
`// split the right atom of ${p}_{Y}$` - 6:
- ${y}^{(3)}=\{{y}_{2},{y}_{2}+{h}_{y},{x}_{2}+2{h}_{y},\dots ,{y}_{2}+({N}_{y}-3){h}_{y}\}$; ${p}_{Y}^{(3)}(y)={\sum}_{i=1}^{{N}_{y}-2}\delta (y-{y}_{i}^{(3)}){p}_{Y}^{(3)}({y}_{i}^{(3)})$
- 7:
- ${h}_{{s}^{\u2605}}=\frac{{x}_{{N}_{x}}+{y}_{{N}_{y}}-({x}_{1}+{y}_{1})}{{N}_{{s}^{\u2605}}-1}$
`// corresponding step size of the main part of convolution grid s` - 8:
- ${h}_{s}=max({h}_{x},{h}_{y},{h}_{{s}^{\u2605}})$
`// final step size of the main part of convolution grid s` - 9:
- ${N}_{s}=\lfloor \frac{{x}_{{N}_{x}}+{y}_{{N}_{y}}-({x}_{1}+{y}_{1})}{{h}_{s}}\rfloor $
`// corresponding number of points` - 10:
`// set irregular convolution grid:`- 11:
**if**${h}_{s}\ge ({x}_{2}+{y}_{2}-{x}_{1}-{y}_{1})$**then**- 12:
- $s=\{{x}_{1}+{y}_{1},{x}_{2}+{y}_{2},{x}_{2}+{y}_{2}+{h}_{s},\dots ,{x}_{2}+{y}_{2}+({N}_{s}-3){h}_{s},{x}_{{N}_{x}}+{y}_{{N}_{y}}\}$
- 13:
**else**- 14:
- $s=\{{x}_{1}+{y}_{1},{x}_{2}+{y}_{2}-{h}_{s},{x}_{2}+{y}_{2},\dots ,{x}_{2}+{y}_{2}+({N}_{s}-4){h}_{s},{x}_{{N}_{x}}+{y}_{{N}_{y}}\}$
- 15:
**end if**- 16:
- ${x}^{{(3)}^{\prime}}=\{{x}_{2},{x}_{2}+{h}_{s},{x}_{2}+2{h}_{s},\dots ,{x}_{2}+({N}_{x}-3){h}_{s}\}$
`// discretize ${x}^{(3)}$ with ${h}_{s}$` - 17:
- ${p}_{X}^{{(3)}^{\prime}}(x)={\sum}_{i=1}^{{N}_{x}^{\prime}}\delta (x-{x}_{i}^{{(3)}^{\prime}}){p}_{X}^{{(3)}^{\prime}}({x}_{i}^{{(3)}^{\prime}})$
`// regrid ${p}_{X}^{(3)}$` - 18:
- ${y}^{{(3)}^{\prime}}=\{{y}_{2},{y}_{2}+{h}_{s},{y}_{2}+2{h}_{s},\dots ,{y}_{2}+({N}_{y}-3){h}_{s}\}$
`// discretize ${y}^{(3)}$ with ${h}_{s}$` - 19:
- ${p}_{Y}^{{(3)}^{\prime}}(y)={\sum}_{i=1}^{{N}_{y}^{\prime}}\delta (y-{y}_{i}^{{(3)}^{\prime}}){p}_{Y}^{{(3)}^{\prime}}({y}_{i}^{{(3)}^{\prime}})$
`// regrid ${p}_{Y}^{(3)}$` - 20:
- ${\mathcal{B}}^{(1)}={p}_{X}^{(1)}\oplus {p}_{Y}^{(1)}$
`// Brute Force convolution` - 21:
- ${\mathcal{B}}^{(2)}={p}_{X}^{(2)}\oplus {p}_{Y}^{(1)}$
`// —”—` - 22:
- ${\mathcal{B}}^{(3)}={p}_{X}^{{(3)}^{\prime}}\oplus {p}_{Y}^{(1)}$
`// —”—` - 23:
- ${\mathcal{B}}^{(4)}={p}_{X}^{(1)}\oplus {p}_{Y}^{(2)}$
`// —”—` - 24:
- ${\mathcal{B}}^{(5)}={p}_{X}^{(2)}\oplus {p}_{Y}^{(2)}$
`// —”—` - 25:
- ${\mathcal{B}}^{(6)}={p}_{X}^{{(3)}^{\prime}}\oplus {p}_{Y}^{(2)}$
`// —”—` - 26:
- ${\mathcal{B}}^{(7)}={p}_{X}^{(1)}\oplus {p}_{Y}^{{(3)}^{\prime}}$
`// —”—` - 27:
- ${\mathcal{B}}^{(8)}={p}_{X}^{(2)}\oplus {p}_{Y}^{{(3)}^{\prime}}$
`// —”—` - 28:
- ${\mathcal{B}}^{(9)}={p}_{X}^{{(3)}^{\prime}}\oplus {p}_{X}^{{(3)}^{\prime}}$
`// —”—` - 29:
- Regrid ${\mathcal{B}}^{(1-9)}$ onto convolution grid s
Output: Discrete probability density function ${p}_{{S}^{\perp}}$ of the independent sum ${S}^{\perp}=X+Y$ with the support$s=\{{s}_{1},{s}_{2},{s}_{2}+{h}_{s},{s}_{2}+2{h}_{s},\dots ,{s}_{2}+({N}_{s}-3){h}_{s},{s}_{{N}_{s}}\}$ and the associated probabilities ${p}_{{S}^{\perp}}(s)={\sum}_{k=1}^{{N}_{s}}\delta (s-{s}_{k}){p}_{{S}^{\perp}}({s}_{k})$, where ${s}_{{N}_{s}}-[{s}_{2}+({N}_{s}-3){h}_{s}]\le {h}_{s}$, ${s}_{2}-{s}_{1}\le {h}_{s}$, ${h}_{s}\ge max({h}_{1},{h}_{2})$. |

Algorithm 2: Brute force convolution for supports with the same span |

Input: Two discrete probability density functions ${p}_{X}$ and ${p}_{Y}$, where the supports of X and Y aredefined using the same span h as: $x=\{{x}_{1},{x}_{1}+h,\dots ,{x}_{1}+({N}_{x}-1)h\}$, $y=\{{y}_{1},{y}_{1}+h,\dots ,{y}_{1}+({N}_{y}-1)h\}$ and the associated probabilities as ${p}_{X}(x)={\sum}_{i=1}^{{N}_{x}}\delta (x-{x}_{i}){p}_{X}({x}_{i})$, ${p}_{Y}(y)={\sum}_{j=1}^{{N}_{y}}\delta (y-{y}_{i}){p}_{Y}({y}_{j})$ - 1:
- $s=\{{x}_{1}+{y}_{1},{x}_{1}+{y}_{1}+h,{x}_{1}+{y}_{1}+2h,\dots ,{x}_{1}+{y}_{1}+({N}_{x}+{N}_{y}-2)h\}$
`// compute convolution support` - 2:
- ${p}_{{S}^{\perp}}(s)=\mathbf{0}$
`// initialize` - 3:
- ${p}_{{S}^{\perp}}({s}_{i+j-1})\leftarrow {p}_{{S}^{\perp}}({s}_{i+j-1})+{p}_{X}({x}_{i}){p}_{Y}({y}_{j})\phantom{\rule{1.em}{0ex}}1\le i\le {N}_{x},1\le j\le {N}_{y}$
`// compute probabilities`
Output: Discrete probability density function ${p}_{{S}^{\perp}}$ of the independent sum ${S}^{\perp}=X+Y$ with the support$s=\{{s}_{1},{s}_{2},\dots ,{s}_{{N}_{x}+{N}_{y}-1}\}$ and the corresponding probabilities ${p}_{{S}^{\perp}}(s)={\sum}_{k=1}^{{N}_{x}+{N}_{y}-1}\delta ({s}_{k}){p}_{{S}^{\perp}}({s}_{k})$. |

#### 2.4.2. Regriding

Algorithm 3: Linear regriding |

Algorithm 4: 4-point regridding, Stage I |

Algorithm 5: 4-point regridding, Stage II |

#### 2.4.3. Comonotonization and Mixture Approximation

Algorithm 6: Distribution of the comonotonic sum |

#### 2.5. Order of Convolutions and Tree Topology

## 3. Results

## 4. Conclusions

## 5. Future Research Directions

## Author Contributions

## Funding

## Conflicts of Interest

## Appendix A

Algorithm A1: Split-Atom Convolution: 4-products |

Input: Two discrete pdfs ${p}_{X}$ and ${p}_{Y}$ with supports:$x=\{{x}_{1},{x}_{1}+{h}_{x},{x}_{1}+2{h}_{x},\dots ,{x}_{1}+({N}_{x}-2){h}_{x},{x}_{{N}_{x}}\}$; $y=\{{y}_{1},{y}_{1}+{h}_{y},{y}_{1}+2{h}_{y},\dots ,{y}_{1}+({N}_{y}-2){h}_{y},{y}_{{N}_{y}}\}$; and probabilities: ${p}_{X}(x)={\sum}_{i=1}^{{N}_{x}}\delta (x-{x}_{i}){p}_{X}({x}_{i})$; ${p}_{Y}(y)={\sum}_{i=1}^{{N}_{y}}\delta (y-{y}_{i}){p}_{Y}({y}_{i})$; ${N}_{s}$ - maximum number of points for discretizing convolution grid - 1:
- ${x}^{(1)}=\{{x}_{1},{x}_{1}+{h}_{x},{x}_{1}+2{h}_{x},\dots ,{x}_{1}+({N}_{x}-2){h}_{x}\}$; ${p}_{X}^{(1)}(x)={\sum}_{i=1}^{{N}_{x}-2}\delta (x-{x}_{i}^{(1)}){p}_{X}^{(1)}({x}_{i}^{(1)})$
- 2:
- ${x}^{(2)}=\{{x}_{{N}_{x}}\}$; ${p}_{X}^{(2)}(x)=\delta (x-{x}_{i}^{(2)})\xb7{p}_{X}^{(2)}({x}_{i}^{(2)})$
`// split the right atom of ${p}_{X}$` - 3:
- ${y}^{(1)}=\{{y}_{1},{y}_{1}+{h}_{y},{y}_{1}+2{h}_{y},\dots ,{y}_{1}+({N}_{y}-2){h}_{y}\}$; ${p}_{Y}^{(1)}(y)={\sum}_{i=1}^{{N}_{y}-2}\delta (y-{y}_{i}^{(1)}){p}_{Y}^{(1)}({y}_{i}^{(1)})$
- 4:
- ${y}^{(2)}=\{{y}_{{N}_{y}}\}$; ${p}_{Y}^{(2)}(y)=\delta (y-{y}_{i}^{(2)})\xb7{p}_{Y}^{(2)}({y}_{i}^{(2)})$
`// split the right atom of ${p}_{Y}$` - 5:
- ${h}_{{s}^{\u2605}}=\frac{{x}_{{N}_{x}}+{y}_{{N}_{y}}-({x}_{1}+{y}_{1})}{{N}_{{s}^{\u2605}}-1}$
`// corresponding step size of the main part of convolution grid s` - 6:
- ${h}_{s}=max({h}_{x},{h}_{y},{h}_{{s}^{\u2605}})$
`// final step size of the main part of convolution grid s` - 7:
- ${N}_{s}=\lceil \frac{{x}_{{N}_{x}}+{y}_{{N}_{y}}-{x}_{1}-{y}_{1}}{{h}_{s}}\rceil +1$
`// corresponding number of points` - 8:
`// set irregular convolution grid:`- 9:
- $s=\{{x}_{1}+{y}_{1},{x}_{1}+{y}_{1}+{h}_{s},\dots ,{x}_{1}+{y}_{1}+({N}_{s}-2){h}_{s},{x}_{{N}_{x}}+{y}_{{N}_{y}}\}$
- 10:
- ${N}_{{x}^{{(1)}^{\prime}}}=\lceil \frac{{x}_{1}+({N}_{x}-2){h}_{x}-{x}_{1}}{{h}_{s}}\rceil +1$
`// size of the regridded main part` - 11:
- ${x}^{{(1)}^{\prime}}=\{{x}_{1},{x}_{1}+{h}_{s},{x}_{1}+2{h}_{s},\dots ,{x}_{1}+({N}_{{x}^{{(1)}^{\prime}}}-1){h}_{s}\}$
`// discretize ${x}^{(1)}$ with ${h}_{s}$` - 12:
- ${p}_{X}^{{(1)}^{\prime}}(x)={\sum}_{i=1}^{{N}_{x}^{\prime}}\delta (x-{x}_{i}^{{(1)}^{\prime}}){p}_{X}^{{(1)}^{\prime}}({x}_{i}^{{(1)}^{\prime}})$
`// regrid ${p}_{X}^{(1)}$` - 13:
- ${N}_{{y}^{{(1)}^{\prime}}}=\lceil \frac{{y}_{1}+({N}_{y}-2){h}_{y}-{y}_{1}}{{h}_{s}}\rceil +1$
- 14:
- ${y}^{{(1)}^{\prime}}=\{{y}_{1},{y}_{1}+{h}_{s},{y}_{1}+2{h}_{s},\dots ,{y}_{1}+({N}_{{y}^{{(1)}^{\prime}}}-1){h}_{s}\}$
- 15:
- ${p}_{Y}^{{(1)}^{\prime}}(y)={\sum}_{i=1}^{{N}_{y}^{\prime}}\delta (y-{y}_{i}^{{(1)}^{\prime}}){p}_{Y}^{{(1)}^{\prime}}({y}_{i}^{{(1)}^{\prime}})$
- 16:
- ${\mathcal{B}}^{(1)}={p}_{X}^{{(1)}^{\prime}}\oplus {p}_{Y}^{{(1)}^{\prime}}$
`// Brute Force convolution` - 17:
- ${\mathcal{B}}^{(2)}={p}_{X}^{(2)}\oplus {p}_{Y}^{{(1)}^{\prime}}$
`// —”—` - 18:
- ${\mathcal{B}}^{(3)}={p}_{X}^{{(1)}^{\prime}}\oplus {p}_{Y}^{(2)}$
`// —”—` - 19:
- ${\mathcal{B}}^{(4)}={p}_{X}^{(2)}\oplus {p}_{Y}^{(2)}$
`// —”—` - 20:
- Regrid ${\mathcal{B}}^{(1-4)}$ onto convolution grid s
Output: Probability mass function ${p}_{S}$ of the independent sum $S=X+Y$ with the support$s=\{{s}_{1},{s}_{1}+{h}_{s},{s}_{1}+2{h}_{s},\dots ,{s}_{1}+({N}_{s}-2){h}_{s},{s}_{{N}_{s}}\}$ and the associated probabilities ${p}_{S}(s)={\sum}_{k=1}^{{N}_{s}}\delta (s-{s}_{k}){p}_{S}({s}_{k})$, where ${s}_{{N}_{s}}-[{s}_{1}+({N}_{s}-2){h}_{s}]\le {h}_{s}$. |

Algorithm A2: Modified local moment matching |

Input: Discrete pdf ${p}_{X}$ with fine scale support $x=\{{x}_{1},{x}_{1}+h,\dots ,{x}_{1}+(N-1)h\}$ and associated probabilities ${p}_{X}(x)={\sum}_{i=1}^{N}\delta (x-{x}_{1}-(i-1)h)\xb7{p}_{X}({x}_{1}+(i-1)h)$; the support of coarse scale probability mass function ${p}_{{X}^{\prime}}$: ${x}^{\prime}=\{{x}_{1}^{\prime},{x}_{1}^{\prime}+{h}^{\prime},\dots ,{x}_{1}^{\prime}+({N}^{\prime}-1){h}^{\prime}\}$. Requirement: ${x}_{1}^{\prime}\le {x}_{1}$, ${x}_{{N}^{\prime}}^{\prime}\ge {x}_{N}$, $h<{h}^{\prime}$.
- 1:
- For convenience, let $\{{x}_{1},{x}_{2},\dots ,{x}_{N}\}\leftarrow x$ and $\{{x}_{1}^{\prime},{x}_{2}^{\prime},\dots ,{x}_{{N}^{\prime}}^{\prime}\}\leftarrow {x}^{\prime}$
- 2:
- Initialize ${p}_{{X}^{\prime}}({x}^{\prime})\leftarrow \mathbf{0}$
- 3:
**if**${x}_{1}^{\prime}={x}_{1}$**then**- 4:
- ${p}_{{X}^{\prime}}({x}_{1}^{\prime})\leftarrow {p}_{{X}^{\prime}}({x}_{1}^{\prime})+{p}_{X}({x}_{1})$
- 5:
**end if**- 6:
**for**$i<{N}^{\prime}$, $j\le N$, ${x}_{i}^{\prime}<{x}_{j}\le {x}_{i+2}^{\prime}$, ${x}_{j}-{x}_{i}^{\prime}\le {x}_{i+3}^{\prime}-{x}_{j}$**do**- 7:
- $$\left[\begin{array}{c}{p}_{{X}^{\prime}}({x}_{i}^{\prime})\\ {p}_{{X}^{\prime}}({x}_{i+1}^{\prime})\\ {p}_{{X}^{\prime}}({x}_{i+2}^{\prime})\end{array}\right]\leftarrow \left[\begin{array}{c}{p}_{{X}^{\prime}}({x}_{i}^{\prime})\\ {p}_{{X}^{\prime}}({x}_{i+1}^{\prime})\\ {p}_{{X}^{\prime}}({x}_{i+2}^{\prime})\end{array}\right]+{\left[\begin{array}{ccc}1& 1& 1\\ {x}_{i}^{\prime}& {x}_{i+1}^{\prime}& {x}_{i+2}^{\prime}{{x}^{\prime}}_{i}^{2}\\ {{x}^{\prime}}_{i}^{2}& {{x}^{\prime}}_{i+1}^{2}& {{x}^{\prime}}_{i+2}^{2}\end{array}\right]}^{-1}\left[\begin{array}{c}{p}_{X}({x}_{j})\\ {x}_{j}{p}_{X}({x}_{j})\\ {x}_{j}^{2}{p}_{X}({x}_{j}^{2})\end{array}\right]$$
- 8:
**end for**
Output: Pdf ${p}_{{X}^{\prime}}({x}^{\prime})$ |

## Appendix B

## References

- AIR-Worldwide. 2015. AIR Hurricane Model for the United States. Available online: https://www.air-worldwide.com/publications/brochures/documents/air-hurricane-model-for-the-united-states-brochure (accessed on 5 May 2019).
- Arbenz, Philipp, Christoph Hummel, and Georg Mainik. 2012. Copula based hierarchical risk aggregation through sample reordering. Insurance: Mathematics and Economics 51: 122–33. [Google Scholar] [CrossRef]
- Artzner, Philippe, Freddy Delbaen, Jean-Marc Eber, and David Heath. 1999. Coherent measures of risk. Mathematical Finance 9: 203–28. [Google Scholar] [CrossRef]
- Chaubey, Yogendra P., Jose Garrido, and Sonia Trudeau. 1998. On the computation of aggregate claims distributions: Some new approximations. Insurance: Mathematics and Economics 23: 215–30. [Google Scholar] [CrossRef]
- Cherubini, Umberto, Elisa Luciano, and Walter Vecchiato. 2004. Copula Methods in Finance. Wiley Finance Series; Chichester: John Wiley and Sons Ltd. [Google Scholar]
- Clark, K. 2015. Catastrophe Risk; International Actuarial Association/Association Actuarielle Internationale. Available online: http://www.actuaries.org/LIBRARY/Papers/RiskBookChapters/IAA_Risk_Book_Iceberg_Cover_and_ToC_5May2016.pdf (accessed on 5 May 2019).
- Cormen, Thomas H., Charles E. Leiserson, Ronald L. Rivest, and Clifford Stein. 2009. Introduction to Algorithms, 3rd ed. Cambridge: MIT Press. [Google Scholar]
- Côté, Marie-Pier, and Christian Genest. 2015. A copula-based risk aggregation model. The Canadian Journal of Statistics 43: 60–81. [Google Scholar] [CrossRef]
- Denuit, Michel, Jan Dhaene, and Carmen Ribas. 2001. Does positive dependence between individual risks increase stop-loss premiums? Insurance: Mathematics and Economics 28: 305–8. [Google Scholar] [CrossRef]
- Dhaene, Jan, Michel Denuit, Marc J. Goovaerts, Rob Kaas, and David Vyncke. 2002. The concept of comonotonicity in actuarial science and finance: Theory. Insurance: Mathematics and Economics 31: 3–33. [Google Scholar] [CrossRef]
- Dhaene, Jan, Daniël Linders, Wim Schoutens, and David Vyncke. 2014. A multivariate dependence measure for aggregating risks. Journal of Computational and Applied Mathematics 263: 78–87. [Google Scholar] [CrossRef]
- Einarsson, Baldvin, Rafał Wójcik, and Jayanta Guin. 2016. Using intraclass correlation coefficients to quantify spatial variability of catastrophe model errors. Paper Presented at the 22nd International Conference on Computational Statistics (COMPSTAT 2016), Oviedo, Spain, August 23–26; Available online: http://www.compstat2016.org/docs/COMPSTAT2016_proceedings.pdf (accessed on 5 May 2019).
- Evans, Diane L., and Lawrence M. Leemis. 2004. Algorithms for computing the distributions of sums of discrete random variables. Mathematical and Computer Modelling 40: 1429–52. [Google Scholar] [CrossRef]
- Galsserman, Paul. 2004. Monte Carlo Methods in Financial Engineering. New York: Springer. [Google Scholar]
- Gerber, Hans U. 1982. On the numerical evaluation of the distribution of aggregate claims and its stop-loss premiums. Insurance: Mathematics and Economics 1: 13–18. [Google Scholar] [CrossRef]
- Grossi, Patricia, Howard Kunreuther, and Don Windeler. 2005. An introduction to catastrophe models and insurance. In Catastrophe Modeling: A New Approach to Managing Risk. Edited by Patricia Grossi and Howard Kunreuther. Huebner International Series on Risk. Insurance and Economic Security. Boston: Springer Science+Business Media. [Google Scholar]
- Hennessy, John L., and David A. Patterson. 2007. Computer Architecture: A Quantitative Approach, 4th ed. San Francisco: Morgan Kaufmann. [Google Scholar]
- Hürlimann, Werner. 2001. Analytical evaluation of economic risk capital for portfolios of gamma risks. ASTIN Bulletin 31: 107–22. [Google Scholar] [CrossRef]
- Iman, Ronald L., and William-Jay Conover. 1982. A distribution-free approach to inducing rank order correlation among input variables. Communications in Statistics-Simulation and Computation 11: 311–34. [Google Scholar] [CrossRef]
- Koch, Inge, and Ann De Schepper. 2006. The Comonotonicity Coefficient: A New Measure of Positive Dependence in a Multivariate Setting. Available online: https://ideas.repec.org/p/ant/wpaper/2006030.html (accessed on 5 May 2019).
- Koch, Inge, and Ann De Schepper. 2011. Measuring comonotonicity in m-dimensional vectors. ASTIN Bulletin 41: 191–213. [Google Scholar]
- Latchman, Shane. 2010. Quantifying the Risk of Natural Catastrophes. Available online: http://understandinguncertainty.org/node/622 (accessed on 5 May 2019).
- Lee, Woojoo, and Jae Youn Ahn. 2014. On the multidimensional extension of countermonotonicity and its applications. Insurance: Mathematics and Economics 56: 68–79. [Google Scholar] [CrossRef]
- Lee, Woojoo, Ka Chun Cheung, and Jae Youn Ahn. 2017. Multivariate countermonotonicity and the minimal copulas. Journal of Computational and Applied Mathematics 317: 589–602. [Google Scholar] [CrossRef]
- McKay, Michael D., Richard J. Beckman, and William J. Conover. 1979. A comparison of three methods for selecting values of input variables in the analysis of output from a computer code. Technometrics 21: 239–45. [Google Scholar]
- McNeil, Alexander J., and Johanna Nešlehová. 2009. Multivariate Archimedean copulas, d-monotone functions and l1–norm symmetric distributions. The Annals of Statistics 37: 3059–97. [Google Scholar] [CrossRef]
- Nelsen, Roger B. 2006. An Introduction to Copulas. Springer Series in Statistics; New York: Springer. [Google Scholar]
- Panjer, Harry H., and B. W. Lutek. 1983. Practical aspects of stop-loss calculations. Insurance: Mathematics and Economics 2: 159–77. [Google Scholar] [CrossRef]
- Robertson, John. 1992. The computation of aggregate loss distributions. Proceedings of the Casualty Actuarial Society 79: 57–133. [Google Scholar]
- Shevchenko, Pavel V. 2010. Calculation of aggregate loss distributions. The Journal of Operational Risk 5: 3–40. [Google Scholar] [CrossRef]
- Venter, Gary G. 2001. D.e. Papush and g.s. Patrik and f. Podgaits. CAS Forum. Available online: http://www.casact.org/pubs/forum/01wforum/01wf175.pdf (accessed on 5 May 2019).
- Venter, Gary G. 2013. Effects of Parameters of Transformed Beta Distributions. CAS Forum. Available online: https://www.casact.org/pubs/forum/03wforum/03wf629c.pdf (accessed on 5 May 2019).
- Vilar, José L. 2000. Arithmetization of distributions and linear goal programming. Insurance: Mathematics and Economics 27: 113–22. [Google Scholar] [CrossRef]
- Walhin, J. F., and J. Paris. 1998. On the use of equispaced discrete distributions. ASTIN Bulletin 28: 241–55. [Google Scholar] [CrossRef]
- Wang, Shaun. 1998. Aggregation of correlated risk portfolios: Models and algorithms. Proceedings of the Casualty Actuarial Society 85: 848–939. [Google Scholar]
- Wojcik, Rafał, Charlie Wusuo Liu, and Jayanta Guin. 2016. Split-Atom convolution for probabilistic aggregation of catastrophe losses. In Proceedings of 51st Actuarial Research Conference (ARCH 2017.1). SOA Education and Research Section. Available online: https://www.soa.org/research/arch/2017/arch-2017-iss1-guin-liu-wojcik.pdf (accessed on 5 May 2019).
- Zhang, Jilian, Kyriakos Mouratidis, and HweeHwa Pang. 2011. Heuristic algorithms for balanced multi-way number partitioning. Paper Presented at the International Joint Conference on Artificial Intelligence (IJCAI), Barcelona, Spain, July 16–22; Research Collection School Of Information Systems. Menlo Park: AAAI Press, pp. 693–98. [Google Scholar]

**Figure 1.**Aggregation of five risks using copula trees. Direct model (upper panel), hierarchical model with sequential topology (middle panel) and hierarchical model with closest pair topology (lower panel). The leaf nodes represent the risks whose aggregate we are interested in. The branching nodes of direct tree represent a multivariate copula model for the incoming individual risks while the branching nodes of hierarchical trees represent a bivariate copula model for the incoming pairs of individual and/or cumulative risks.

**Figure 2.**Illustration of hypothetical correlation matrices: (

**A**) exchangeable, (

**B**) nested block diagonal, and (

**C**) unstructured correlation matrix.

**Figure 3.**A discrete loss pdf represented as a mixture of two “spikes” (atoms) at minimum and maximum x damage ratio (red) and the main part (blue). Damage ratio is discretized on 64-point grid.

**Figure 4.**An example of RNN approach for determining topology of hierarchical risk aggregation tree for six risks with zero minima. The maxima and cumulative maxima characterizing losses for the six risks are presented in the upper panel. (

**A**) The algorithm takes the largest cumulative max and halves it to obtain the number c. Then, it binary searches for the number closest to c except for the last element in the sequence. This number (showed in bold) becomes the cumulative maximum of the new subsequence. The search is repeated until the subsequence consists of two elements. (

**B**) The resulting hierarchical aggregation tree.

**Figure 5.**An example of recursive nearest neighbor (RNN) approach for determining topology of direct risk aggregation tree for six risks shown in Figure 4. Note that the order of comonotonic aggregation follows the order of independent aggregation.

**Figure 6.**MC (red line) and convolution/comonotoinization based (blue bars) distributions of the total risk for 29,139 locations affected by hurricane peril using different aggregation models with linear regriding (upper row) and 4-point regriding (lower row). No tail truncation was applied. For consistency, the losses are plotted in [0; $100 MM] interval.

**Figure 7.**(

**A**–

**E**) Percentage errors in statistics of the total risk relative to the corresponding values obtained from MC simulations. Risk aggregation was performed for 29,139 locations affected by hurricane peril using sequential (blue) and RNN (red) models with 4-point regriding and maximum support size varying from 64 to 6400; (

**F**) shows the average time cost of five runs for each maximum support size.

**Table 1.**Mean ($\mu $), standard deviation ($\sigma $) and tail Value-at-Risk (TVaR

_{%}) at levels 1%, 5% and 10% of the total hurricane risk for 29,139 locations for (A) direct, (B) hierarchical sequential and (C) hierarchical RNN aggregation models compared to the average values from 30 realizations of MC runs. The losses are in [$MM]. Numbers in brackets represent percentage errors relative to MC simulations.

MC | Linear Regriding No Truncation | Linear Regriding Tail Truncation | 4-Point Regridding | ||
---|---|---|---|---|---|

(A) | $\mu $ | 44.3 | 44.3 (0.00%) | 44.3 (0.00%) | 44.3 (0.00%) |

$\sigma $ | 7.1 | 10.9 (53.02%) | 7.1 (0.00%) | 7.1 (0.00%) | |

TVaR_{1%} | 51.4 | 2550.9 (4864.2%) | 2550.8 (4864.0%) | 2550.8 (4863.9%) | |

TVaR_{5%} | 48.5 | 502.1 (934.7%) | 501.4 (932.8%) | 501.9 (933.8%) | |

TVaR_{10%} | 47.6 | 252.0 (429.7%) | 250.7 (427.1%) | 251.0 (427.6%) | |

(B) | $\mu $ | 44.3 | 44.3 (0.00%) | 44.3 (0.00%) | 44.3 (0.00%) |

$\sigma $ | 7.1 | 7.6 (6.11%) | 7.1 (0.00%) | 7.1 (0.00%) | |

TVaR_{1%} | 93.7 | 81.2 (−13.3%) | 84.7 (−9.6%) | 88.0 (−6.1%) | |

TVaR_{5%} | 66.3 | 62.5 (−5.8%) | 68.5 (3.4%) | 66.8 (0.8%) | |

TVaR_{10%} | 58.6 | 67.3 (15.0%) | 60.5 (3.3%) | 59.8 (2.1%) | |

(C) | $\mu $ | 44.3 | 44.3 (0.00%) | 44.3 (0.00%) | 44.3 (0.00%) |

$\sigma $ | 7.1 | 7.2 (1.55%) | 7.1 (0.00%) | 7.1 (0.00%) | |

TVaR_{1%} | 75.9 | 151.0 (98.9%) | 95.0 (25.2%) | 76.0 (0.1%) | |

TVaR_{5%} | 62.9 | 184.4 (193.4%) | 98.0 (55.8%) | 63.5 (0.9%) | |

TVaR_{10%} | 58.4 | 95.8 (63.9%) | 87.0 (48.9%) | 59.1 (1.1%) |

**Table 2.**Processing times [s] for different risk aggregation models. MC run is a single realization with 1 MM samples. The mixture method implementation for hierarchical trees was optimized for performance with nested block diagonal correlation structure in Figure 2B. Intel i7-4770 CPU @ 3.40 GHz architecture with 16 GB RAM was used.

Aggregation Model | MC | Linear Regridding No Truncation | Linear Regriding Tail Truncation | 4-Point Regriding |
---|---|---|---|---|

Direct | 1539 | 0.25 | 0.26 | 0.33 |

Hierarchical, sequential | 12,769 | 0.34 | 0.35 | 0.44 |

Hierarchical, RNN | 10,512 | 0.40 | 0.41 | 0.52 |

© 2019 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).

## Share and Cite

**MDPI and ACS Style**

Wójcik, R.; Liu, C.W.; Guin, J. Direct and Hierarchical Models for Aggregating Spatially Dependent Catastrophe Risks. *Risks* **2019**, *7*, 54.
https://doi.org/10.3390/risks7020054

**AMA Style**

Wójcik R, Liu CW, Guin J. Direct and Hierarchical Models for Aggregating Spatially Dependent Catastrophe Risks. *Risks*. 2019; 7(2):54.
https://doi.org/10.3390/risks7020054

**Chicago/Turabian Style**

Wójcik, Rafał, Charlie Wusuo Liu, and Jayanta Guin. 2019. "Direct and Hierarchical Models for Aggregating Spatially Dependent Catastrophe Risks" *Risks* 7, no. 2: 54.
https://doi.org/10.3390/risks7020054