Customer Loyalty Analytics

By Dr Mohamed Zaki

Customer loyalty is a strategic priority for organisations. We did some work with one of our industrial partners to build a data-driven method to better assess and predict customer loyalty. Organisations still use single-question customer metrics, such as the Net Promoter Score (NPS), which is popular despite recent studies arguing that customer loyalty is multidimensional, and therefore firms require to combine behavioural and attitudinal data sources. One of the reasons why organisations rely on NPS is the simplicity to administer measuring customer loyalty. However, the picture now gets complicated because customers interact with firms that are leveraging new technologies such as mobile applications, social media platforms, virtual reality, drones and the Internet of Things to provide smart services and enable a seamless customer experience. The complexity of using these technologies within an organisation’s myriad touchpoints has led to a data explosion across touchpoints in the entire customer journey. Thus, it is more difficult to rely on single metrics like the NPS. In the Cambridge Service Alliance, we investigated this area and built a novel customer loyalty analytics method that demonstrates an approach to utilising data more effectively to assess and predict customer loyalty in complex business-to-business (B2B) service organisations.

To acquire a holistic view of customer loyalty, we integrated data across multiple systems. The data was classified into three categories: attitudinal, behavioural and demographics. The attitudinal data was collected from the customer survey, which includes structured (NPS rating) and unstructured data (verbatim comments). The behavioural data was collected from the financial system. This data consists of sales (new, used, lease), product support (parts and service transaction types) and customer service agreement (CSA) transactions (parts and service transaction types). Two groups of customers were identified: those who have a maintenance contract with the company, referred to as Customer Service Agreement (CSA) customers; and those who deal in a transactional setting, referred to as Product Support (PS) customers. Demographic data, which contains the regional locations of customers, was included. In total, we collected 1,044,512 transaction records over a three-year period.

Our predictive model used:

1. RFM to transform customer transactional data into profitability scores, facilitating the categorisation of customers based on their purchasing behaviour
2. The K-means clustering model technique to segment data points into groups, each containing data points similar to one another and dissimilar to data points in other groups. In our work, we divided customers into 11 groups based on their RFM scores. This was accomplished using the K-means segmentation algorithm
3. Active Customers Predictive-We have identified active customers who had not churned and had company dealings in the form of transactions
4. Demographic- we used the geographical location of customers, which appears within a transaction record.
5. Text-mining Model- We developed a linguistic-based text-mining model to analyse the open-ended customer comments in the survey. A sentiment score for each comment was then calculated.

Our customer loyalty model enabled us to classify customers as either churners or loyal customers based on these predictive indicators. Our predictive model was built using neural and Bayesian network classification techniques. The accuracy results of these two algorithms were compared. The steps employed in the model’s construction are three-fold. First, a training set, 60 per cent of the entire data, was used to develop a training model. Then we tested the model against new data in the construction stage, which formed 30 per cent of the original data. Here the model was fine-tuned to decrease the error of false predictions. Finally, the model was validated against the remaining 10 per cent of customer data. These three steps are performed to ensure the repeatability and validity of the prediction model.

Our model shows how firms can compare NPS scores with repurchasing behaviour as a loyalty assessor, using predictive variables such as the RFM model, demographics, active customers and textual customer complaints, while the most popular performance measure, the NPS, completely disregards this. Clearly, customer loyalty measurement requires holistic loyalty-tracking initiatives. Based on the amalgamation of data sources, the actual underlying customer loyalty can be fully assessed and interpreted through the use of big data analytics. Furthermore, our model has the predictive ability to determine whether customers are likely to churn, thereby increasing the model’s functionality. Over the three years, if the organisation used NPS as an indicator for customer loyalty, over half of the customers were considered to be completely satisfied. However, using our analysis, we identified many misclassified NPS categories, which has misled the company. For example, in 2013 and 2014 we found that approximately 500 customers were considered to be detractors when they were classified as promoters according to the NPS classification. Thus, NPS alone is not sufficiently accurate for organisations. If organisations want to understand why their customers churn, the answer could come from the verbatim comments provided in survey data or social media. Our text-mining model enables us to analyse the root causes of customer complaints, which, expressed in these free verbatim comments, uncover potentially vulnerable customers that NPS would have considered loyal and not requiring intervention strategies.

If you want to read more about this work, please read our paper or listen to my webinar