Why Predictive Analytics is More Than Just Good Data
Predictive analytics applies mathematical techniques to convert data into useful information, such as predictions about your customers. This certainly requires good data and advanced models – both are essential – however, the process involves a lot more than just that...
This useful information may predict the next product each customer is likely to purchase, identify probable future churners, or assign customers to segments with different interests and needs. Or it may be another outcome that can help your business to manage customers more efficiently.
The key thing is that predictive analytics is designed to create information that is correctly calculated and beneficial to your business. The calculation of model scores can be complex and there are many ways for errors to creep in - but only one way to compute them correctly. Likewise, there is only one way to measure the real benefit: by testing and experimentation.
In order to achieve success and avoid the risk of failure, predictive analytics involves a series of stages, from problem definition through to testing business benefits. In many ways, the process is like a journey: you start by planning where you intend to go, assess the available modes of transport, travel towards your destination, and finally, arrive.
In predictive analytics, the initial planning stage is known as business requirements. This stage recognizes that a complete understanding of the precise goal is needed in order to create new information that is both useable, i.e. can be applied by your business, and potentially useful, i.e. delivers a worthwhile improvement over current methods. If an essential requirement is overlooked, then the project will be at risk – in the sense that the solution cannot be applied for some reason.
For example, if your board required a set of 5 customer segments in order to monitor customer behaviour, then it would be no good if 10 segments were delivered instead. The more detailed results might be more accurate, however, they would never be used.
The next stages of the journey concern data – to understand your existing sources of customer data, select which sources should be considered, and transform them into a useful format for further analysis. The majority of the total development effort goes into this latter step alone, however, it cannot be skimped – the predictive power of your data is the main driver of success and is more important than employing the most powerful advanced modelling technique.
Analysis and modelling come next, which is where your analyst builds a model that predicts the required outcome while also making sound business sense.
The analyst will assess the effectiveness of the model, by measuring its discriminatory power and comparing the results for various subsets of customers. The acid test comes at the deployment stage.
The first deployment should be a controlled experiment, in which some customers receive a marketing action targeted by the model, while the remainder continues to receive the standard treatment that you normally employ. These targeted and control groups are then monitored to observe their reactions and compare their performance, based on the key indicators of success for this activity.
By designing a live test on the performance of your model, you are able to measure the effectiveness of predictive analytics and be confident that its implementation in your business process will be worthwhile.
For this reason, successful use of predictive analytics requires more than just good data and advanced models – and the process doesn’t stop there! ‘Best practice’ users of predictive analytics continuously employ testing to measure the benefit of their models and obtain the earliest indications of change in their performance.
