# Individual Loss Reserving Using a Gradient Boosting-Based Approach

^{*}

^{†}

## Abstract

**:**

## 1. Introduction and Motivation

## 2. Loss Reserving

- If ${T}_{1}^{(k)}<{t}^{*}<{T}_{2}^{(k)}$, the accident has happened but has not yet been reported to the insurer. It is therefore called an “incurred but not reported” (IBNR), claim. For one of those claims, the insurer does not have specific information about the accident, but can use policyholder and external information to estimate the reserve.
- If ${T}_{2}^{(k)}<{t}^{*}<{T}_{3}^{(k)}$, the accident has been reported to the insurer but is still not settled, which means the insurer expects to make additional payments to the insured. It is therefore called a “reported but not settled” (RBNS), claim. For one such claim, the historical information as well as policyholder and external information can be used to estimate the reserve.
- If ${t}^{*}>{T}_{3}^{(k)}$, the claim is classified as settled, or S, and the insurer does not expect to make more payments.

## 3. Models for Loss Reserving

#### 3.1. Bootstrap Mack’s Model and Generalized Linear Models for Loss Reserving

#### 3.2. Gradient Boosting for Loss Reserving

`R`programming language in conjunction with

`caret`and

`xgboost`libraries.

`caret`is a powerful package used to train and to validate a wide range of statistical models including XGBoost algorithm.

- 1.
- The simplest solution is to train the model on data ${\mathcal{D}}_{\mathcal{T}}$ where only settled claims (or non-censored claims) are included. Hence, the response is known for all claims. However, this leads to a selection bias because claims that are already settled at ${t}^{*}$ tend to have shorter developments, and claims with shorter development tend to have lower total paid amounts. Consequently, the model is almost exclusively trained on simple claims with low training responses, which leads to underestimation of the total paid amounts for new claims. Furthermore, a significant proportion of the claims are removed from the analysis, which causes a loss of information. We will analyze this bias further in Section 4 (see model B).
- 2.
- In Lopez et al. (2016), a different and interesting approach is proposed: in order to correct the selection bias induced by the presence of censored data, a strategy called “inverse probability of censoring weighting” (IPCW) is implemented, which involves assigning weights to observations to offset the lack of complete observations in the sample. The weights are determined using the Kaplan-Meier estimator of the censoring distribution, and a modified CART algorithm is used to make the predictions.
- 3.
- A third approach is to develop claims that are still open at ${t}^{*}$ using parameters from a classical approach such as Mack’s or the GLM model. We discuss this idea in more detail in Section 4 (see model C and model D).

- we can assume that they are static, which can lead to a bias in the predictions obtained (see model E in Section 4); or
- we can, for each of these variables, (1) adjust a dynamic model, (2) obtain a prediction of the complete trajectory, and (3) use the algorithm conditionally to the realization of this trajectory. Moreover, there may be dependence between these variables, which would warrant a multivariate approach.

## 4. Analyses

#### 4.1. Data

#### 4.2. Training of XGBoost Models

#### 4.3. Learning of Prediction Function

Algorithm 1: Obtaining ${\widehat{f}}_{A}$ with least square TreeBoost. |

#### 4.4. Results

- Mack’s model, for which we present results obtained with the bootstrap approach developed by England and Verrall (2002), based on both quasi-Poisson and gamma distributions; and
- generalized linear models for which we present results obtained using a logarithmic link function and a variance function $\mathcal{V}(\mu )=\varphi {\mu}^{p}$ with $p=1$ (quasi-Poisson), $p=2$ (gamma), and $1<p<2$ (Tweedie).

- individual generalized linear models (see Section 3.1), for which we present results obtained using a logarithmic link function and three variance functions: $\mathcal{V}(\mu )=\mu $ (Poisson) and $\mathcal{V}(\mu )=\varphi {\mu}^{p}$ with $p=1$ (quasi-Poisson) and $\mathcal{V}(\mu )=\varphi {\mu}^{p}$ with $1<p<2$ (Tweedie); and
- XGBoost models (models A, B, C, D and E) described in Section 4.2.

## 5. Conclusions

- (1)
- the censored nature of the data could strongly bias the loss reserving process; and
- (2)
- the use of a micro-level model based solely on generalized linear models could be unstable for loss reserving but an approach combining a macro-level (or a micro-level) model for the artificial completion of open claims and a micro-level gradient-boosting model represents an interesting approach for an insurance company.

## Author Contributions

## Funding

## Acknowledgments

## Conflicts of Interest

## Appendix A. Structure of the Dataset

Policy Number | Claim Number | Party | File Number | Date | … |
---|---|---|---|---|---|

P100000 | C234534 | 1 | F0000001 | 31 March 2004 | … |

P100000 | C234534 | 1 | F0000001 | 30 June 2004 | … |

P100000 | C234534 | 1 | F0000001 | 30 September 2004 | … |

… | … | … | … | … | … |

P100000 | C234534 | 2 | F0000002 | 31 March 2004 | … |

P100000 | C234534 | 2 | F0000002 | 30 June 2004 | … |

P100000 | C234534 | 2 | F0000002 | 30 September 2004 | … |

… | … | … | … | … | … |

P100034 | C563454 | 1 | F0000140 | 31 March 2004 | … |

P100034 | C563454 | 1 | F0000140 | 30 June 2004 | … |

P100034 | C563454 | 1 | F0000140 | 30 September 2004 | … |

… | … | … | … | … | … |

## References

- Antonio, Katrien, and Richard Plat. 2014. Micro-level stochastic loss reserving for general insurance. Scandinavian Actuarial Journal 7: 649–69. [Google Scholar] [CrossRef]
- Arjas, Elja. 1989. The claims reserving problem in non-life insurance: Some structural ideas. ASTIN Bulletin 19: 140–52. [Google Scholar] [CrossRef]
- Baudry, Maximilien, and Christian Y. Robert. 2017. Non Parametric Individual Claim Reserving in Insurance. Working paper. [Google Scholar]
- Breiman, Leo, Friedman Jerome, Olshen Richard, and Stone Charles. 1984. Classification and Regression Trees. Wadsworth Statistics/Probability Series; New York: Routledge. [Google Scholar]
- Buhlmann, Hans, Rene Schnieper, and Erwin Straub. 1980. Claims reserves in casualty insurance based on a probabilistic model. Bulletin of the Association of Swiss Actuaries 80: 21–45. [Google Scholar]
- Charpentier, Arthur, and Mathieu Pigeon. 2016. Macro vs. micro methods in non-life claims reserving (an econometric perspective). Risks 4: 12. [Google Scholar] [CrossRef]
- Chen, Tianqi, and Carlos Guestrin. 2016. Xgboost: A scalable tree boosting system. Paper presented at 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, San Francisco, CA, USA, August 13–17; pp. 785–94. [Google Scholar]
- De Jong, Piet, and Gillian Z. Heller. 2008. Generalized Linear Models for Insurance Data. Cambridge: Cambridge University Press. [Google Scholar]
- England, Peter D., and Richard J. Verrall. 2002. Stochastic claims reserving in general insurance. British Actuarial Journal 8: 443–544. [Google Scholar] [CrossRef]
- Friedman, Jerome H. 2001. Greedy function approximation: A gradient boosting machine. Annals of Statistics 29: 1189–232. [Google Scholar] [CrossRef]
- Gabrielli, Andrea, Ronald Richman, and Mario V. Wüthrich. 2019. Neural Network Embedding of the Over-Dispersed Poisson Reserving Model. Scandinavian Actuarial Journal 2019: 1–29. [Google Scholar] [CrossRef]
- Haastrup, Svend, and Elja Arjas. 1996. Claims reserving in continuous time; a nonparametric bayesian approach. ASTIN Bulletin 26: 139–64. [Google Scholar] [CrossRef]
- Hachemeister, Charles. 1980. A stochastic model for loss reserving. Transactions of the 21st International Congress of Actuaries 1: 185–94. [Google Scholar]
- Hesselager, Ole. 1994. A Markov model for loss reserving. ASTIN Bulletin 24: 183–93. [Google Scholar] [CrossRef]
- Hiabu, Munir, Margraf Carolin, Martínez-Miranda Maria, and Nielsen Jens Perch. 2016. The link between classical reserving and granular reserving through double chain ladder and its extensions. British Actuarial Journal 21: 97–116. [Google Scholar] [CrossRef]
- Huang, Jinlong, Chunjuan Qiu, and Xianyi Wu. 2015. Stochastic loss reserving in discrete time: Individual vs. aggregate data models. Communications in Statistics—Theory and Methods 44: 2180–206. [Google Scholar] [CrossRef]
- Jewell, William S. 1989. Predicting IBNYR events and delays. ASTIN Bulletin 19: 25–55. [Google Scholar] [CrossRef]
- Larsen, Christian Roholte. 2007. An individual claims reserving model. ASTIN Bulletin 37: 113–32. [Google Scholar] [CrossRef]
- Lopez, Olivier, Xavier Milhaud, and Pierre-E. Thérond. 2016. Tree-based censored regression with applications in insurance. Electronic Journal of Statistics 10: 2685–716. [Google Scholar] [CrossRef]
- Mack, Thomas. 1993. Distribution-free calculation of the standard error of chain ladder reserve estimates. ASTIN Bulletin: The Journal of the IAA 23: 213–25. [Google Scholar] [CrossRef]
- Norberg, Ragnar. 1986. A contribution to modeling of IBNR claims. Scandinavian Actuarial Journal 1986: 155–203. [Google Scholar] [CrossRef]
- Norberg, Ragnar. 1993. Prediction of outstanding liabilities in non-life insurance. ASTIN Bulletin 23: 95–115. [Google Scholar] [CrossRef]
- Norberg, Ragnar. 1999. Prediction of outstanding liabilities. II Model variations and extensions. ASTIN Bulletin 29: 5–25. [Google Scholar] [CrossRef]
- Pigeon, Mathieu, Katrien Antonio, and Michel Denuit. 2013. Individual loss reserving with the multivariate skew normal framework. ASTIN Bulletin 43: 399–428. [Google Scholar] [CrossRef]
- Taylor, Greg, Gráinne McGuire, and James Sullivan. 2008. Individual claim loss reserving conditioned by case estimates. Annals of Actuarial Science 3: 215–56. [Google Scholar] [CrossRef]
- Wüthrich, Mario V. 2018. Machine learning in individual claims reserving. Scandinavian Actuarial Journal. in press. [Google Scholar] [CrossRef]
- Wüthrich, Mario V., and Michael Merz. 2008. Stochastic Claims Reserving Methods in Insurance. Zürich and Tübigen: Wiley. [Google Scholar]
- Zhao, Xiaobing, and Xian Zhou. 2010. Applying copula models to individual claim loss reserving methods. Insurance: Mathematics and Economics 46: 290–99. [Google Scholar] [CrossRef]
- Zhao, Xiaobing, Xian Zhou, and Jing Long Wang. 2009. Semiparametric model for prediction of individual claim loss reserving. Insurance: Mathematics and Economics 45: 1–8. [Google Scholar] [CrossRef]

**Figure 2.**(

**a**) Status of claims of incurred amounts on 31 December 2016; (

**b**) Means and standard deviations of incurred amounts on 31 December 2016.

**Figure 3.**Comparison of predictive distributions (incurred but not reported (IBNR) + reported but not settled (RBNS)) for collective models.

Model | Response Variable (${\widehat{\mathit{C}}}_{{\mathit{T}}_{3}}$) | Covariates | Usable in Practice? |
---|---|---|---|

Model A | $\{{C}_{{T}_{3}}\}$ | ${\mathit{x}}_{7}$ | No |

Model B | ${\{{C}_{7}^{(k)}\}}_{k\in {\mathcal{T}}_{B}}$, ${\mathcal{T}}_{B}=\{k\in \mathcal{T}:{T}_{3}^{(k)}<7\}$ | ${\mathit{x}}_{7}$ | Yes |

Model C | closed: $\{{C}_{7}\}$ | ${\mathit{x}}_{7}$ | Yes |

open: $\{{\widehat{\lambda}}_{j}^{c}{C}_{7}\}$ ($\widehat{\lambda}$ from bootstrap) | ${\mathit{x}}_{7}$ | ||

Model D | all: $\{{C}_{7}+{\sum}_{t=8}^{13}{\widehat{Y}}_{t}\}$ (with ${\widehat{Y}}_{t}={q}_{{\mathit{Y}}_{t}}({\kappa}_{D})$) | ${\mathit{x}}_{7}$ | Yes |

Model E | closed: $\{{C}_{7}\}$ | ${\mathit{x}}_{13}$ | No |

open: $\{{\widehat{\lambda}}_{j}^{c}{C}_{7}\}$ ($\widehat{\lambda}$ from bootstrap) | ${\mathit{x}}_{13}$ |

Development Year | 1 | 2 | 3 | 4 | 5 | 6 | 7 |
---|---|---|---|---|---|---|---|

Accident year | |||||||

2004 | 79 | 102 | 66 | 49 | 57 | 48 | 37 |

2005 | 83 | 128 | 84 | 55 | 52 | 41 | · |

2006 | 91 | 138 | 69 | 49 | 38 | · | · |

2007 | 111 | 155 | 98 | 61 | · | · | · |

2008 | 100 | 178 | 99 | · | · | · | · |

2009 | 137 | 251 | · | · | · | · | · |

2010 | 155 | · | · | · | · | · | · |

Development Year | 1 | 2 | 3 | 4 | 5 | 6 | 7 | $8+$ |
---|---|---|---|---|---|---|---|---|

Accident year | ||||||||

2004 | 34 | 41 | 23 | 13 | 14 | 14 | 9 | 7 |

2005 | 37 | 60 | 36 | 29 | 45 | 21 | 20 | 24 |

2006 | 41 | 64 | 34 | 23 | 21 | 14 | 4 | 21 |

2007 | 46 | 67 | 40 | 37 | 15 | 18 | 3 | 13 |

2008 | 46 | 82 | 39 | 42 | 16 | 11 | 15 | 33 |

2009 | 54 | 109 | 62 | 51 | 31 | 36 | 11 | 2 |

2010 | 66 | 93 | 47 | 45 | 16 | 16 | 9 | ? |

**Table 4.**Prediction results (incurred but not reported (IBNR) + reported but not settled (RBNS)) for collective approaches.

Model | Assessment | $\mathbf{E}\phantom{\rule{-0.166667em}{0ex}}\left[\mathbf{Res}.\right]$ | $\sqrt{\mathbf{Var}\phantom{\rule{-0.166667em}{0ex}}\left[\mathbf{Res}.\right]}$ | ${\mathit{q}}_{0.95}$ | ${\mathit{q}}_{0.99}$ |
---|---|---|---|---|---|

Bootstrap Mack | out-of-sample | 76,795,136 | 7,080,826 | 89,086,213 | 95,063,184 |

(quasi-Poisson) | in-sample | 75,019,768 | 8,830,631 | 90,242,398 | 97,954,554 |

Bootstrap Mack | out-of-sample | 76,803,753 | 7,170,529 | 89,133,141 | 95,269,308 |

(Gamma) | in-sample | 75,004,053 | 8,842,412 | 90,500,323 | 98,371,607 |

GLM | out-of-sample | 75,706,046 | 2,969,877 | 80,655,890 | 82,696,002 |

(Quasi-Poisson) | in-sample | 74,778,091 | 3,084,216 | 79,922,183 | 81,996,425 |

GLM | out-of-sample | 73,518,411 | 2,263,714 | 77,276,416 | 78,907,812 |

(Gamma) | in-sample | 71,277,218 | 3,595,958 | 77,343,035 | 80,204,504 |

GLM | out-of-sample | 75,688,916 | 2,205,003 | 79,317,520 | 80,871,729 |

(Tweedie) | in-sample | 74,706,050 | 2,197,659 | 78,260,722 | 79,790,056 |

Model | Assessment | $\mathbf{E}\phantom{\rule{-0.166667em}{0ex}}\left[\mathbf{Res}.\right]$ | $\sqrt{\mathbf{Var}\phantom{\rule{-0.166667em}{0ex}}\left[\mathbf{Res}.\right]}$ | ${\mathit{q}}_{0.95}$ | ${\mathit{q}}_{0.99}$ | |
---|---|---|---|---|---|---|

Poisson | out-of-sample | 86,411,734 | 9007 | 86,426,520 | 86,431,211 | |

in-sample | 75,611,203 | 8655 | 75,625,348 | 75,631,190 | ||

Quasi-Poisson | out-of-sample | 86,379,296 | 894,853 | 87,815,685 | 88,309,697 | |

in-sample | 75,606,230 | 814,608 | 76,984,768 | 77,433,248 | ||

Tweedie | out-of-sample | 84,693,529 | 2,119,280 | 88,135,187 | 90,011,542 | |

in-sample | 70,906,225 | 1,994,004 | 74,098,686 | 75,851,991 |

Model | $\mathbf{E}\phantom{\rule{-0.166667em}{0ex}}\left[\mathbf{Res}.\right]$ | $\sqrt{\mathbf{Var}\phantom{\rule{-0.166667em}{0ex}}\left[\mathbf{Res}.\right]}$ | ${\mathit{q}}_{0.95}$ | ${\mathit{q}}_{0.99}$ |
---|---|---|---|---|

Model A | 73,204,299 | 3,742,971 | 79,329,916 | 82,453,032 |

Model B | 14,339,470 | 6,723,608 | 25,757,061 | 30,643,369 |

Model C | 67,655,960 | 2,411,739 | 71,708,313 | 73,762,242 |

Model D | 68,313,731 | 4,176,418 | 75,408,868 | 78,517,966 |

Model E | 67,772,822 | 2,387,476 | 71,722,744 | 73,540,516 |

© 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**

Duval, F.; Pigeon, M. Individual Loss Reserving Using a Gradient Boosting-Based Approach. *Risks* **2019**, *7*, 79.
https://doi.org/10.3390/risks7030079

**AMA Style**

Duval F, Pigeon M. Individual Loss Reserving Using a Gradient Boosting-Based Approach. *Risks*. 2019; 7(3):79.
https://doi.org/10.3390/risks7030079

**Chicago/Turabian Style**

Duval, Francis, and Mathieu Pigeon. 2019. "Individual Loss Reserving Using a Gradient Boosting-Based Approach" *Risks* 7, no. 3: 79.
https://doi.org/10.3390/risks7030079