| Name: | Description: | Size: | Format: | |
|---|---|---|---|---|
| 2.43 MB | Adobe PDF |
Authors
Advisor(s)
Abstract(s)
Este trabalho propõe novos métodos para inferir a competitividade das apólices de seguro automóvel quando a informação disponível sobre preços é limitada. Funcionalidades do espaço dos espaços são utilizadas para filtrar o ruído das observações, introduzindo dependências temporais para a transição entre seguradoras ou para conversão. Foram recolhidos dados sobre transições entre companhias de seguros no mercado português para estimar as probabilidades de transição entre seguradoras. O modelo binomial oculto de Markov mostrou-se algo limitado ao pressupor um espaço de estados discreto. O filtro de Kalman foi mais bem sucedido na remoção do ruído das observações. O filtro de Kalman proporcionou resultados intuitivos que são interpretáveis mesmo para um público não técnico. Também se utilizaram dados de conversão para inferir estimativas semanais de alteração de competitividade. Propusemos modelos de regressão penalizada em que o tempo é incluído como uma estrutura de passeio aleatório. O modelo utiliza ponderadores de credibilidade para combinar alterações em cada segmento com as alterações da carteira. A estrutura hierárquica do modelo produz estimativas para as alterações de competitividade que são mais interpretáveis do que as dos modelos lineares generalizados, onde o tempo é incluído como uma variável categórica. Além disso, o método proposto supera os modelos lineares generalizados em termos de desempenho preditivo. Ambos os métodos podem servir como uma ferramenta para apoiar o processo de tomada de decisão sobre preços por parte das seguradoras, quando a disponibilidade de informação fiável sobre preços é limitada.
This work proposes several novel methods for inferring competitiveness of motor insurance policies in a setting of limited availability of price information. State-space functionalities are employed to filter noise from observations by introducing underlying time-dependent structures for transition and conversion data. Transition data of insurance companies of vehicles in the Portuguese insurance market was collected to analyze the evolution of the incoming transition probabilities of insurers. The binomial hidden Markov model is somewhat restricted due to its assumption of discrete state-space. The Kalman smoother is more successful in removing noise from the observations. The smoother provides intuitive results that are interpretable for a non-technical audience. Furthermore, conversion data was used to infer weekly segment-specific estimates of competitiveness changes. We have proposed a penalized regression framework where time is included as a random walk structure. The model uses credibility weighting on each segment's changes using the full portfolio's changes as the complement. The powerful hierarchical fashion of the model produces estimates of competitiveness changes that are more interpretable than those of generalized linear models, where time is included as a categorical variable. Moreover, the proposed method outperforms the generalized linear models in terms of predictive performance. Both methods can serve as a tool to support the price decision-making process by insurers when the availability of reliable price information is limited.
This work proposes several novel methods for inferring competitiveness of motor insurance policies in a setting of limited availability of price information. State-space functionalities are employed to filter noise from observations by introducing underlying time-dependent structures for transition and conversion data. Transition data of insurance companies of vehicles in the Portuguese insurance market was collected to analyze the evolution of the incoming transition probabilities of insurers. The binomial hidden Markov model is somewhat restricted due to its assumption of discrete state-space. The Kalman smoother is more successful in removing noise from the observations. The smoother provides intuitive results that are interpretable for a non-technical audience. Furthermore, conversion data was used to infer weekly segment-specific estimates of competitiveness changes. We have proposed a penalized regression framework where time is included as a random walk structure. The model uses credibility weighting on each segment's changes using the full portfolio's changes as the complement. The powerful hierarchical fashion of the model produces estimates of competitiveness changes that are more interpretable than those of generalized linear models, where time is included as a categorical variable. Moreover, the proposed method outperforms the generalized linear models in terms of predictive performance. Both methods can serve as a tool to support the price decision-making process by insurers when the availability of reliable price information is limited.
Description
Mestrado em Actuarial Science
Keywords
Seguro automóvel competitividade modelo oculto de Markov filtro de Kalman regressão penalizada Motor insurance competitiveness hidden Markov model Kalman filter penalized regression
Pedagogical Context
Citation
Meindersma, Johannes (2020). "Inferring competitiveness without price information : an application to the motor insurance portfolio of Fidelidade". Dissertação de Mestrado. Universidade de Lisboa. Instituto Superior de Economia e Gestão.
Publisher
Instituto Superior de Economia e Gestão
