A carregar...
Publicação
Do GLM ao LightGBM: inovação na previsão de anulação de apólices automóvel
| Nome: | Descrição: | Tamanho: | Formato: | |
|---|---|---|---|---|
| 2.89 MB | Adobe PDF |
Orientador(es)
Resumo(s)
Num contexto empresarial de elevada competitividade, a retenção de clientes constitui um fator estratégico essencial para a sustentabilidade das seguradoras. Este estudo, desenvolvido em parceria com a Allianz Portugal, tem como objetivo construir um modelo de previsão de anulação de apólices de seguro automóvel com recurso a técnicas de Machine Learning. A investigação procura aperfeiçoar os modelos preditivos existentes, explorando alternativas ao tradicional Generalized Linear Model (GLM), amplamente utilizado no setor segurador pela sua robustez e interpretabilidade. A metodologia baseia-se na comparação entre dois modelos: o GLM, desenvolvido no software Emblem, e o modelo baseado em árvores de decisão com boosting, implementado através do algoritmo LightGBM, em Python. O foco recai na antecipação de padrões de risco de anulação,
cuja previsão precisa impacta a eficiência operacional e a retenção de clientes. O estudo utilizou uma amostra de dados históricos da companhia, submetida a análise exploratória, limpeza de inconsistências e redução da multicolinearidade, garantindo a qualidade dos dados. A modelação recorreu à otimização automática de hiperparâmetros com a biblioteca Optuna e a validação cruzada, assegurando robustez e generalização dos resultados. A métrica de avaliação adotada foi a Area Under the Curve (AUC), adequada a problemas de classificação binária. Os resultados indicam que o LightGBM supera o GLM em capacidade discriminativa, evidenciando ganhos consistentes na previsão de anulações. Esta investigação reforça o papel da ciência de dados na modernização analítica do setor segurador e destaca o potencial das técnicas avançadas de machine learning em contextos empresariais reais.
In an increasingly competitive business environment, customer retention plays a strategic role in ensuring the sustainability of insurance companies. This study, developed in collaboration with Allianz Portugal, aims to develop a predictive model for the cancellation of motor insurance policies using machine learning techniques. The research seeks to enhance the existing predictive models, exploring alternatives to the traditional Generalized Linear Model (GLM), widely employed in the insurance sector for its robustness and interpretability. The methodology is based on a comparison between two predictive approaches: the GLM,developed in Emblem software, and a boosted decision tree model implemented using the LightGBM algorithm in Python. The analysis focuses on anticipating patterns associated with cancellation risk, as accurate forecasting of this behavior directly affects operational efficiency and customer retention. The study employed a historical dataset provided by the insurance company, which underwent exploratory data analysis (EDA), data cleaning, and multicollinearity reduction to ensure data quality. Model development included automatic hyperparameter optimization using the Optuna library and cross-validation to guarantee robustness and generalization of results. The evaluation metric adopted was the Area Under the Curve (AUC), particularly suited to binary classification problems. The results demonstrate that the LightGBM model outperforms the GLM in terms of discriminative capability, showing consistent improvements in the prediction of policy cancellations. This research highlights the role of data science in modernizing analytical processes within the insurance industry and emphasizes the potential of advanced machine learning techniques in real-world business contexts.
In an increasingly competitive business environment, customer retention plays a strategic role in ensuring the sustainability of insurance companies. This study, developed in collaboration with Allianz Portugal, aims to develop a predictive model for the cancellation of motor insurance policies using machine learning techniques. The research seeks to enhance the existing predictive models, exploring alternatives to the traditional Generalized Linear Model (GLM), widely employed in the insurance sector for its robustness and interpretability. The methodology is based on a comparison between two predictive approaches: the GLM,developed in Emblem software, and a boosted decision tree model implemented using the LightGBM algorithm in Python. The analysis focuses on anticipating patterns associated with cancellation risk, as accurate forecasting of this behavior directly affects operational efficiency and customer retention. The study employed a historical dataset provided by the insurance company, which underwent exploratory data analysis (EDA), data cleaning, and multicollinearity reduction to ensure data quality. Model development included automatic hyperparameter optimization using the Optuna library and cross-validation to guarantee robustness and generalization of results. The evaluation metric adopted was the Area Under the Curve (AUC), particularly suited to binary classification problems. The results demonstrate that the LightGBM model outperforms the GLM in terms of discriminative capability, showing consistent improvements in the prediction of policy cancellations. This research highlights the role of data science in modernizing analytical processes within the insurance industry and emphasizes the potential of advanced machine learning techniques in real-world business contexts.
Descrição
Palavras-chave
Machine Learning GLM LightGBM Seguro Automóvel Anulação Apólices Motor Insurance Policy Cancellation