6 New Technology Drivers reducing Churn
- Time management
- Target management
- Models as features
- Rewards, not regrets
- Feature importance
- AI-enabled UX
New methods for determining forecast/execution windows aid greatly in dealing with customers who churn. Customers who churn become more focused over time, in one case, 52% of were in the top decile 90 days out, vs 71% 30 days out. Customer events drive over 60% of the insights into why they churn, and in many cases, are extremely time-sensitive. Timeframe ( hours/days/weeks ) & forecasting trajectory over time really do matter.
Churn is a multi-customer event process tracked in multiple systems: CRM, ERP, support, operations, etc. Some events are “hard” churn events ( “port-out” ), other “soft” ( “sentiment” to leave via support call ), some inbound, others outbound. Deciding what churn events & time frame 30/60/90 days to target is no longer a one-size-fits-all decision. Multiple targets can now be scored & adjudicated using OR techniques.
Models as Features
In our omnichannel world, traditional methods put all CX variables in one model but disparities in signal strength, timing, the nature of inputs, all can confuse ML models. Open-source modeling technology may not solve these issues. State-of-the-art churn modeling takes into account other CX issues such as future spend on products & services, upsell, cross-sell, etc., all such scoring feeding churn models as features.
Rewards, not Regrets
Reinforcement learning (RL) is the most potent way to learn what actions & offers ( “rewards” ) succeed in stopping a customer from churning. Traditional RL brings “regrets” as some customers will churn because of wrong actions. State-of-the-art application of RL technology uses retroactive & counter-factual methods to develop first policies to jump-start the learning process to arrive at optimum NBAs with minimal “regrets”.
Traditional methods in assigning weights to features fail to manage continuous & categorical variables. For maximum benefit, we must organize features & derivatives according to stems & relevant business ontologies – then map the relative contributions of each. Methods to understand feature importance such as by Shapley ( 2012 Nobel Prize Economics ) are now required methods for any CX targeting, especially churn.
Customers are too important to be left to any single team – marketing, sales, support, etc., Likewise, reducing churn is way too important to be left exclusively to data scientists. CX systems must operate in digital channels in near real-time, enabling personalized actions & offers ( “NBAs” ). Measuring & scoring churn is not enough, the same systems must also make it easy to use the latest in related NBAs.
To everything ( churn, churn, churn ) There is a season ( churn, churn, churn ) ( apologies to The Byrds & their 1965 hit song Turn! Turn! Turn! )
The importance of managing churn increases as we turn to digital e-commerce and as the cost of acquiring new customers increases relative to the cost to keep customers.
At its core, churn is an evolutionary process for the user and a discrete process for the vendor, a unique challenge.
In either B2C or B2B scenarios, a customer may decide to churn over time ( port-out a phone line in wireless, scale back their number of users in a subscription plan, cancel their entire subscription plan, etc. ). B2C & B2B customers may experience several churn-inducing events in their journeys, such as bad service, competitive alternatives, and better pricing elsewhere.
Many such events can be important in driving customers to churn, yet vendors may be blind to these events until late in the customers’ decision-making. As a result, vendors typically react to events and are often soon thereafter thrust into recovery mode, with is reactive and extremely expensive.
The customary way in which modeling is done to identify customers who might churn is as follows:
Set the target: usually the customer leaving. This customer event is typically captured in a vendor’s ERP system and possibly in their CRM or billing system.
Collect data set: from a representative subset of the base that includes churners and non-churners alike.
Extract dataset: then clean, organize, and transform. Generate features that are extracted from the data set and load into EDL or equivalent.
Simple statistics: perform simple statistical tests ( variance, t-student ) checking to see if some features in the data allow us to distinguish customers who churn from those who do not churn ( non-churners ).
Model: choose a modeling technique, which usually means using open source libraries. The most commonly used technique is decision trees: the dataset is classified by continuously dividing the “tree” until reaching a final “decision”. Decision trees are very understandable and are common to many rules-based decision engine designs. Once a modeling technique is chosen, then the entire configuration is coded accordingly.
Test: run the churn model developed through standard testing & validation, evaluating performance in terms of “lift” provided by the model.
Deploy: the final version of the model for enterprise production.
Cerebri AI churn modeling:
Cerebri AI technology leverages three main techniques for churn modeling:
- Engineered Features ( EFs )
- Multi-target/ Multi-insight Models
- Advanced Modeling Techniques
Over the course of the past five years, Cerebri AI has developed and organized engineered features including mathematical transformations, timing & covariation transformations, business transformations, causality, sentiment analysis, & targeted models, all designed to help tease the signal for CX targets from the data noise in customer journeys