| Nome: | Descrição: | Tamanho: | Formato: | |
|---|---|---|---|---|
| 6.81 MB | Adobe PDF |
Autores
Orientador(es)
Resumo(s)
O mercado de telecomunicações e actualmente caracterizado por forte concorrência entre os vários operadores em actividade e por um elevado nível de saturação, sendo cada vez mais difícil a angariação de novos clientes, havendo por isso uma forte aposta na retenção e fidelização dos existentes. Por este motivo é cada vez mais importante o conhecimento do perfil do cliente, a identificação dos factores que influenciam a sua satisfação e das variáveis que o influenciam realmente na decisão de mudar de prestador de serviço ou de se manter com o actual. Com o objectivo de caracterizar os clientes que desactivam os seus serviços recolheu-se informação relativa a um segmento específico: clientes residenciais pós-pagos. A formulação de um (ou vários) modelo(s) de regressão logística - para variável resposta do tipo binário, activo ou desactivo – servirá para identificar quais são os factores que estes clientes mais valorizam e têm real impacto na sua satisfação, bem como, em oposição, identificar claramente quais os podem levar à decisão de mudar de operador, ou seja, encontrar os factores que diferenciam os clientes activos dos desactivos. Os modelos de regressão logística são um caso particular de um vasto conjunto de modelos de utilização muito ampla: os modelos lineares generalizados. Estes caracterizam-se pelo facto de poderem ter variável resposta não normal, desde que esta satisfaça a condição de ser bem ajustada por uma distribuição pertencente à familia exponencial. A ligação entre a variável resposta e o vector de covariáveis pode ser estabelecida através de uma função monótona diferenciável chamada função de ligação. É apresentado neste estudo um método de estimaçao dos parâmetros para este tipo de modelos. Sao descritas várias estratégias de modelação e comparados os resultados respectivos, sendo também descritos alguns problemas numéricos surgidos durante o processo (comuns para dados deste tipo), algumas possíveis causas e soluções.
Telecomunications business is presently marked by fierce competition amongst operators and high saturation level, therefore leading to growing difficulties to acquire new customers. Due to this situation market players are increasingly focusing on retention and loyalty programs to maintain current ones. Strong knowledge of customer profile is gaining great importance, since knowing what are the factors that influence customers’ satisfaction and can make one decide to change service provider (or keep the present one) can be of great help to design retention programs and focus on decisive / really important variables. With the goal of characterizing deactivated customer profile, all available information related with a specific segment – post-paid residentials – has been gathered. Regression models – for binary dependent variables, active or deactivated – were formulated based on this data. These models are aimed to help identify which factors are valued and have a real impact on customers, and to find out which can lead them to the decision of changing to another provider. The purpose is therefore to identify the factors which differentiate active customers form deactivated ones. Logistic regression models are a particular case of a much wider class of models vastly used: generalized linear models. These can have a non linear response variable, as long as it is well approximated by any distribution belonging to the exponential family. The relationship between the dependent variable and the independent ones can be established through a differentiable and monotone function called the link function. An estimation method for the model parameters is presented in this paper. Distinct modeling strategies are described and compared in this study. Some numerical problems (common for this type of data) arisen during the modeling process are also detailed, as well as their possible causes and some solutions.
Telecomunications business is presently marked by fierce competition amongst operators and high saturation level, therefore leading to growing difficulties to acquire new customers. Due to this situation market players are increasingly focusing on retention and loyalty programs to maintain current ones. Strong knowledge of customer profile is gaining great importance, since knowing what are the factors that influence customers’ satisfaction and can make one decide to change service provider (or keep the present one) can be of great help to design retention programs and focus on decisive / really important variables. With the goal of characterizing deactivated customer profile, all available information related with a specific segment – post-paid residentials – has been gathered. Regression models – for binary dependent variables, active or deactivated – were formulated based on this data. These models are aimed to help identify which factors are valued and have a real impact on customers, and to find out which can lead them to the decision of changing to another provider. The purpose is therefore to identify the factors which differentiate active customers form deactivated ones. Logistic regression models are a particular case of a much wider class of models vastly used: generalized linear models. These can have a non linear response variable, as long as it is well approximated by any distribution belonging to the exponential family. The relationship between the dependent variable and the independent ones can be established through a differentiable and monotone function called the link function. An estimation method for the model parameters is presented in this paper. Distinct modeling strategies are described and compared in this study. Some numerical problems (common for this type of data) arisen during the modeling process are also detailed, as well as their possible causes and some solutions.
Descrição
Tese de mestrado, Probabilidades e Estatística, Universidade de Lisboa, Faculdade de Ciências, 2009
Palavras-chave
Telecomunicações Modelos lineares generalizados Regressão logística Resposta binária Teses de mestrado - 2009
