In highly competitive industries, customer retention has received much attention. Customer retention is an important issue, as loyal customers tend to produce greater cash flow and profits, are less sensitive to price, bring along new customers and do not require any acquisition or start-up costs.
Predicting Retention Rates
We have also used neural networks to learn to distinguish insurance policy holders who are likely to terminate their policies from those who are likely to renew in order to predict the retention rate prior to price sensitivity analysis. Policy holders of an Australian motor insurance company are classified into 30 risk groups based on their demographic and policy information using k-means clustering (Yeo, Smith, Willis & Brooks, 2001, 2003). Neural networks are then used to model the effect of premium price change on whether a policy holder will renew or terminate his or her policy. A multilayered feedforward neural network was constructed for each of the clusters with 25 inputs and 1 output (whether the policy holder renews or terminates the contract).
Several experiments were carried out on a few clusters to determine the most appropriate number of hidden neurons and the activation function. Twenty hidden neurons and the hyperbolic tangent activation function were used for the neural networks for all the clusters. A uniform approach is preferred to enable the straight-forward application of the methodology to all clusters, without the need for extensive experimentation by the company in the future. Input variables that were skewed were log transformed.
Some of the issues we encountered in using neural networks to determine the effect of premium price change on whether a policy holder will renew or terminate his or her policy were:
Determining the threshold for classifying policy holders into those who terminate and those who renew
Generating more homogenous models
Clusters that had too few policy holders to train the neural networks
Key Terms in this Chapter
Genetic Programming: Search method inspired by natural selection. The basic idea is to evolve a population of “programs” candidates to the solution of a specific problem.
Weights: Strength of a connection between two neurons in a neural network.
Multi-Layered Feedforward Network: A layered neural network in which each layer only receives inputs from previous layers.
Association Rules: Predict the occurrence of an event based on the occurrences of another event.
Regression Analysis: A statistical technique used to find relationships between variables for the purpose of predicting future variables.
Decision Trees: Tree-shaped structures that represent sets of decisions. These decisions generate rules for the classification of a dataset.
Rough Sets: Rough sets are mathematical algorithms that interpret uncertain, vague, or imprecise information.
Confusion Matrix: Contains information about actual and predicted classifications done by a classification system.
Neuron: The basic processing element of a neural network.
Backpropagation: Method for computing the error gradient for a feedforward neural network.
Neural Network: Non-linear predictive models that learn through training and resemble biological neural networks in structure.
K-Means Clustering: An algorithm that performs disjoint cluster analysis on the basis of Euclidean distances computed from variables and randomly generated initial seeds.
Boosting: Generates multiple models or classifiers (for prediction or classification), and to derive weights to combine the predictions from those models into a single prediction or predicted classification.
Multiple Concept-Level Association Rules: Extend association rules from single level to multiple levels. Database contents are associated together to the concepts, creating different abstraction levels.