Think about how people usually approach analytics. Either they segment data, looking for clues and connecting dots, or they develop key performance indicators (KPIs) in advance and react to what it spits out. Each of these have drawbacks.
With normal segmenting techniques, what I’ll refer to as “slicing and dicing”, you are generally only looking at a single variable at a time. So you might know that people prefer product X, but you won’t know that it is driven largely by men using Chrome after 10pm at night. You’d have to do a tremendous amount of slicing and dicing to figure this out.
With KPIs, you are proactively defining what is important and you’re probably right and will get good data on those KPIs. But you won’t get what you aren’t looking for, such as the insight that product growth is being driven largely by men using Chrome after 10pm at night.
There’s a different way to look at analytics and that is predictively. You ask your data specific questions and it gives specific answers. This is accomplished by feeding your analytics into a predictive model such as the Google Prediction API. You choose the question you want answered and then feed in all the previous answers you got to the same question. Any obvious example is a recommendation engine. The question you are asking is “What product would this person like?” You feed in all the orders from everyone with as much demographic, preference and behavior data as possible. Then you can ask what products people will like based on that same set of demographic, preference and behavior data.
The power of predictive analytics is far greater than recommendations, though. The same prediction model that recommends products to customers can be mined for analytics data. Using the example above, while designing a marketing program, we could ask the same prediction model “What products do women between 30-40 shopping from home at night on IE prefer?” We can also write a script to sweep through examples and look for particularly strong correlations.
The astute reader may ask, “How is this different than slicing and dicing data?” It’s different because the predictive model looks for all the relationships between all of the data at once. It encapsulates away the heavy mathematics that you need to look for multi-variant relationships. Instead of you sitting there segmenting data, the prediction model looks at every segment at once. You can’t answer the question “What products do women between 30-40 shopping from home at night on IE prefer?” with Google Analytics. You can answer it with the Google Prediction API.
But the real beauty of predictive analytics is that you think like you normally think. You write down the questions you want answered. You gather the data you have on how that question was answered in the past. You can then predict the answer for future data as well as mine the model for hidden relationships. It’s the Bayesian approach to analytics. And it’s the future.
Next time: a concrete example. You can help! Take this ridiculously short survey.