Predictive Analytics to Avert Customer Churn
Our customer is a global IT service provider, one of the most recognized companies in the IT community.
Our customer is committed to building productive and long–term client relationships that result in enduring loyalty to the company. Our customer strongly believes that the best way to meet this challenge is a dynamic response to the problems that end consumers may experience when using hardware or software produced by our client.
The project was aimed at creating a predictive (statistical) model that from a number of complaints received daily is able to choose critical ones. Using this model, our customer will be able to respond proactively leveraging additional resources to solve the problem and thus to avoid the occurrence of critical situations.
Within the project, IBA built a statistical (predictive) model designed to prevent customer churn and increase the level of trust to the equipment manufactured by our client.
Any mathematically grounded forecasts are based on the processing of historical data (data sample). To build a model for predicting critical situations with regard to the equipment, IBA analyzed historical complaints data for the last three years. Statistically, less than 10 of 1000 complaints resulted in critical situations. The analyzed data sample included both regular (minor) complaints and complaints that were eventually escalated to critical situations.
The sample contains three levels of data (input variables), namely those related to:
- The problem itself, including its urgency, the time when it was detected, and the time spent to solve it
- The equipment where the problem has occurred, including the frequency of previous complaints for the last six months and their urgency
- The end consumer, whose equipment experienced a problem, including the frequency of complaints and their urgency.
The data in the sample are historical, which means that it is known whether the critical situation has occurred or not. In the final sample, one of the columns reflects the characteristics of the complaint, also called the dependent variable, which should be predicted by the values of other fields (input variables). In this case, the dependent variable adopts two values: 1 corresponds to the critical complaint that results in customer churn, and 0 to the status of a loyal customer.
Based on this sample, a statistical model is built. The algorithm of the model defines which combinations result in 1 and 0 value of dependent variable. Before a statistical model is selected, it is necessary to execute certain data cleansing tasks, namely to eliminate extremes and outliers and to handle the missing data. Depending on a situation, various statistical and data–mining algorithms can be used for model–building, including logistic regression, decision trees, and neural networks. The model is selected based on a “training” subset of the data sample and is verified against a “test” subset of the data sample.
In addition, to improve the accuracy of prediction, IBA used the algorithms of unstructured data (text data) analysis that assess the textual description of the client's primary complaint and highlight key words and phrases.
The predictive (statistical) model built by IBA reveals about 83 percent of all critical complaints well in advance. The model recognizes critical situations and allows for a stand–by period from four to five days so that our customer will be able to react proactively. To date, our client's support team has enough time for a dynamic response to the complaint selected by the model and for preventing a critical situation. Using the IBA–created model, our customer is able to reduce the risk of critical situations significantly and to enhance client loyalty.