Wouldn’t it be great if we can one day predict the brand of car our customers will buy when they are at their peak, what brand of cornflakes they would eat every Sunday morning, when they will watch another 2012/Independence Day-type movie, right down to what gender of baby they are likely to have when they marry? “Predicting the unpredictable, quantifying the unquantifiable”, says Dogbert the guru*.  But before we are able to fully achieve that in the twilight years, lets see how we can predict the future with the pre-historical method of predictive analytics, summarized in eight easy steps (inspired by Dogbert).

1.) Use a previous campaign or a test as your base.

Say for example you need a conversion of 1000 respondents, and the industry/benchmark average is 2%, then we probably need a base of 50,000 to achieve your target or do an adequate test.

2.) Add geographic, demographic, psychographic & behavioural data to your base

Both the responders and the non-responders. Who wants to be known as Segment A when we can be described more adequately and be given a nice personality.

3.) Know what are all the possible ways (and the most effective way) of reaching your base

We do have mobile phones and a life in the social space.

4.) Test and Control

Divide your data into at least 2 groups everytime, equally. Test group carries the hypothesis, control group validates.

5.) Trash the anomalies

Don’t you hate it when you thought one of your content pillar is doing extremely well in terms of average time spent and when you deep dive into the data you realise that it’s some idiot who left the computer on with your website in the browser running throughout the lonely night? Discard the outliers, be it purchase patterns or web patterns.

6.) Design your modeling framework

We will usually start out with a typical multiple regression model before we move on the the more sophisticated models like CHAID etc. “A regression is an equation that describes the relationship between a dependent variable and more than one independant variable”.  Statistical definition of dependant variable connotes the action/consequence that will be influenced by how you set up the test environment, and indenpendant variable as the components that will not be influenced (e.g. demographic, geographic, behavioural data).

7.) Grade and weigh each variable and develop an algorithm

This step will help you in shortlisting the most important variables that will influence your results (e.g. income, age, family size, etc). One will probably need a PHD in statistics and a SAS or SPSS software to run a regression model and weigh the different variables, and to finally develop an algorithm ranking the deciles.

8.) Score the validation group

If the predicted results derived from the algorithm is a close match to your previous campaign/test, then  the algorithm developed will be useful to predict future campaigns. It should also help you identify the target segment more likely to respond to your campaign through the scoring exercise.

If all else fails, hire a consultant to do the work while you make yourself another cuppa! (*highly recommended)

Last but not least, meet Dogbert, the data guru.