|
Title:
|
PREDICTING CUSTOMER ATTRITION WITH MARKOV CHAIN |
|
Author(s):
|
Krzysztof Dzieciolowski |
|
ISBN:
|
978-989-8533-39-5 |
|
Editors:
|
Ajith P. Abraham, Antonio Palma dos Reis and Jörg Roth |
|
Year:
|
2015 |
|
Edition:
|
Single |
|
Keywords:
|
Markov chain, multinomial model, contract, customer attrition |
|
Type:
|
Short Paper |
|
First Page:
|
157 |
|
Last Page:
|
161 |
|
Language:
|
English |
|
Cover:
|
|
|
Full Contents:
|
click to dowload 
|
|
Paper Abstract:
|
Customer attrition, also referred to as churn, occurs when customers terminate their service with the firm. It represents a loss of recurring revenue to the firm and is a main financial problem plaguing many industries such as banking and telecommunications. Customer attrition is used as one of the key measurement of firms performance and monitored periodically to detect and quantify customer turnover so that the corrective actions could be taken.. The paper suggests a novel application of Markov chains to predict customer attrition. A bank-financed loan or mobile company subsidized-handset typically takes a form of a contract of certain duration and conditions of repayment. It is assumed that each month customers may change one of the four states defined as a regular monthly payment, voluntary attrition, renewal of contract and non-payment, with the latter one sometimes leading to an involuntary attrition. The Markov chain model is a natural representation of such a process. The transition probabilities with which customers change their states are usually non-homogeneous since they depend on the number of months left on contract. The fewer months left on the contract, the lower a penalty for breaking it up and, hence, the higher a probability of attrition. We propose a innovative method of estimating the Markov chains transition probabilities using multinomial regression model with time-to-expiration covariates. Distinct monthly models are built to account for a seasonality of transition probabilities. Non-contract and new customers are also considered to make the prediction model complete. The model is illustrated with real data. |
|
|
|
|
|
|
