A carregar...
Publicação
Modelação e Predição do Valor do tempo de vidado cliente (Customer Lifetime Value)
| Nome: | Descrição: | Tamanho: | Formato: | |
|---|---|---|---|---|
| 2.67 MB | Adobe PDF |
Orientador(es)
Resumo(s)
O conceito "valor de tempo de vida do cliente",Customer Lifetime Value(CLV) da literaturaanglo-saxónica, surgiu devido à necessidade de ao longo dos anos, as empresas reteremos clientes mais lucrativos. Conhecendo o valor de cada um dos seus clientes, as empresasconseguem alocar os seus recursos de forma mais consciente, bem como determinar omáximo de investimento que é viável fazer em cada cliente para o conseguir reter.Ao longo dos anos têm sido desenvolvidos vários métodos para a medição do CLV como,por exemplo, as abordagens RFM (Recency, Frequency and Monetetary Value), em que ovalor do cliente é calculado tendo por base as variáveis recência, frequência e valor mo-netário da compra, ou SOW (Share-of-Wallet) que consiste na simples segmentação dosclientes como "bons" ou "maus". Este tipo de soluções, mais antigas, tem como principaldesvantagem o facto de segmentarem os clientes, tendo apenas em conta a contribui-ção passada do cliente. Para ultrapassar este problema alguns autores propuseram, maisrecentemente, a utilização de modelos deMachine Learningeframeworksdebig datacon-seguindo obter uma maior precisão na previsão do CLV.Os objetivos deste estudo são a construção de modelos preditivos do CLV.Neste estudo, numa primeira fase, foi feita a revisão de literatura sobre o conceito deCLV, bem como sobre as técnicas usadas para o calcular e modelar. Seguiu-se a imple-mentação de modelos deMachine Learningnomeadamente dos modelosClassification andRegression Trees(CART),Random Forest(RF), Cadeias de Markov,Multi Layer Perceptron(MLP) e algoritmos declustering. No final fez-se uma comparação entre esses métodos ecaracterização de cada um dos grupos resultante da análise declusters.Odatasetusado provêm de uma instituição bancária portuguesa, incluindo variáveisdemográficas e sócio-económicas. A partir do conjunto inicial de variáveis foi feita aconstrução de novas variáveis, que foi sempre apoiada na revisão da literatura e simulta-neamente ajustada ao negócio bancário.Conclui-se que, as RF são o modelo que apresenta maior eficácia na previsão do CLV,registando um erro percentual de 5.98%.Neste estudo, foi também realizada uma análise declusterscom o propósito de obter ummelhor conhecimento dos produtos que suscitam interesse a cada grupo de clientes e con-sequentemente servir de base para desenhar campanhas demarketingpersonalizadas. Emparticular, foram encontrados 5clustersde clientes interesses e valores de CLV distintos.
The concept of Customer Lifetime Value (CLV) comes from Anglo-Saxon literature andstarted emerging when companies felt the need to retain the most profitable customers.Therefore when a company knows the value of each of their customers, they are able toallocate their resources more effectively and determine the maximum investment that isfeasible to make in each customer to retain it.Over the years various methods have been developed to measure CLV, such as the RFM(Recency, Frequency and Monetary Value) approach, where the customer value is calcu-lated based on the variables recency, frequency, and monetary value, or SOW (Share-of-Wallet) which consists of simply segmenting customers as "good"or "bad". Comparatively,in this older type of solution, the main disadvantage is the fact that they divide customers,taking into account only the past contribution of the customer. In order to overcome thisproblem some authors have proposed, more recently to use Machine learning models andbig data frameworks achieving greater accuracy in CLV forecast.The objectives of this study is the construction of predictive models of CLV.In this study, we start with a literature review on the concept of CLV as well as the tech-niques used to calculate and model it, followed by the implementation of the MachineLearning models, in particular Classification and Regression Trees (CART), RandomForests (RF), Markov Chains, Multi-Layer Perceptron (MLP), and clustering algorithms.Finally, we make a comparison between these methods and characterize each of them intogroups resulting from cluster analysis.The dataset used comes from a Portuguese banking institution and includes demographicand sociology-economic variables, from the initial set of variables we constructed newvariables, which has always been supported in the literature review and simultaneouslyadjusted to the banking business.We concluded the Random Forest (RF) model is the most effective in forecasting the CLV,recording a percentage error of 5.98%. Finally, we analyzed the cluster with the intent ofbetter understanding the product interest of each group, thus building more personalizedmarketing campaigns bearing in mind the CLV value for each group. We found 5 clustersof clients with different CLV values and interests.
The concept of Customer Lifetime Value (CLV) comes from Anglo-Saxon literature andstarted emerging when companies felt the need to retain the most profitable customers.Therefore when a company knows the value of each of their customers, they are able toallocate their resources more effectively and determine the maximum investment that isfeasible to make in each customer to retain it.Over the years various methods have been developed to measure CLV, such as the RFM(Recency, Frequency and Monetary Value) approach, where the customer value is calcu-lated based on the variables recency, frequency, and monetary value, or SOW (Share-of-Wallet) which consists of simply segmenting customers as "good"or "bad". Comparatively,in this older type of solution, the main disadvantage is the fact that they divide customers,taking into account only the past contribution of the customer. In order to overcome thisproblem some authors have proposed, more recently to use Machine learning models andbig data frameworks achieving greater accuracy in CLV forecast.The objectives of this study is the construction of predictive models of CLV.In this study, we start with a literature review on the concept of CLV as well as the tech-niques used to calculate and model it, followed by the implementation of the MachineLearning models, in particular Classification and Regression Trees (CART), RandomForests (RF), Markov Chains, Multi-Layer Perceptron (MLP), and clustering algorithms.Finally, we make a comparison between these methods and characterize each of them intogroups resulting from cluster analysis.The dataset used comes from a Portuguese banking institution and includes demographicand sociology-economic variables, from the initial set of variables we constructed newvariables, which has always been supported in the literature review and simultaneouslyadjusted to the banking business.We concluded the Random Forest (RF) model is the most effective in forecasting the CLV,recording a percentage error of 5.98%. Finally, we analyzed the cluster with the intent ofbetter understanding the product interest of each group, thus building more personalizedmarketing campaigns bearing in mind the CLV value for each group. We found 5 clustersof clients with different CLV values and interests.
Descrição
Palavras-chave
Customer Lifetime Value (CLV) Banca Rentabilidade Machine Learning Clustering Árvores de Decisão