As real business cases demonstrate, the performance of predictive data mining models drastically improve when information about historical customer behavior and real-time interactions are put together. Companies with traditional churn prediction models should realize the opportunities they are missing…
You can download PDF version of this whitepaper here.
Real time business intelligence and data analytics is becoming popular among most leading companies. More and more, focus is being put on the ability to react to customer interactions the moment they occur, rather than later when the real opportunity is lost. Yet, very few blend the benefits of traditional data mining models with real-time analytics. For example, the accuracy of a data mining model that predicts the likelihood of customer churn the next two months can drastically improve when combined with the complaints information during those two months. A red flag should be raised when a customer who is identified to be at risk by a predictive model takes a certain action that is also identified as a customer churn trigger.
In predictive analytics, there is always a balance between how early the model predicts and how recent data it uses as input. The former is a requirement by the business, as it leaves enough time for taking preventive actions. It is also, usually, required as the ETL and scoring processes take too long and can not be repeated frequently over time. The latter, however, defines how accurate a model can be, as the customers usually show their signs of disloyalty when they are just about to leave. The common approach to balancing these two has been coming up with two sets of analysis:
- One for dealing with the real-time churn triggers to handle the recent customer interactions – such as setting warning mechanisms for high-risk complaints.
- Another for predicting customer churn a few weeks – and sometimes months – earlier, using predictive data mining models.
However, these two approaches can be utilized perfectly, to develop data mining models, incorporating the benefits of both. The performance of such mixed models beat both alternatives as standalones, and, in fact, can even multiply their accuracy. A sanitized case demonstrates the impact of mixed-models:
Company X – Middle East – Leading Telecommunications Player
- The company has performed an analysis on the customer complaints and found out that customers making certain complaints have a risk of 7% churn in the same month.
- The company also developed churn prediction models, which identify a relatively high churn risk group, which is likely to churn by 10% after 2 months.
- Putting the churn model and complaints data together for a mixed model yield to identification of the highest churn risk groups, which are likely to churn by 50% after 2 months, if they make certain complaints.
Incorporating the real-time data with longer-term predictive models can save millions of dollars for companies trying to decrease their churn rates, by letting them focus on the best target groups.
We recommend three main steps towards building a mixed model for accurate prediction of churn, cross-sales or any business problem of a similar nature:
- Prepare Data: Unlike regular data mining models, the data required for mixed models include both historical and most recent data. While certain customer behavior dimensions should take into account the state of customer a certain time period back (e.g. customer value, product ownership, etc. 2 months ago), the triggers which should be incorporated should be recent (e.g. complaints last week).
- Perform Preliminary Analysis: The preliminary analysis should be conducted to identify the triggers which are relevant for the business problem. For example, if the company wants to predict churn, complaints which are most critical should be identified analyzing their correlations with customer churn. The preliminary analysis should be also performed to understand the correlations between customer behavior dimensions and churn as well (e.g. customers with short tenure are more likely to churn).
- Build the Mixed Model: The approach to building the mixed predictive model, once the data is ready and selected, is not very different from building a regular data mining model. Variables related to churn, from different timeframes, are put together and fed into any of the predictive modeling techniques that are available (e.g. logistic regression, decision trees, neural networks). The result will be models which identify customers at high risk, utilizing both recent and longer-term profiles (e.g. customer who has decreased usage in last 6 months, has short tenure and made a complaint last week regarding prices).
Once the model is ready, it should be put into practice across channels, where they trigger certain business actions, when the customer interaction leads to a significant increase in churn risk, as scored by the model.
Gaining ability to accurately identify which customers will churn and taking reactive measures to prevent them is an effective means for customer retention. However, for the best solution, companies should focus on a more pro-active approach and develop solutions for the churn triggers identified during the churn prediction analysis.
To receive more information about our recommendations, and learn about related service offerings, please contact: firstname.lastname@example.org.