The Sophistication of Prediction

Two weeks ago I posted about the Marketing AI Landscape, and promised a slightly deeper dive into one of the analytical subject areas. This is that deeper dive, and in it, I want to outline how I see AI as simply a further step on a sophistication path of greater automation and newer algorithms. Improving sophistication is exciting, but shouldn’t necessarily be sought at the expense of the business objective.

I think it helps to focus on an individual analytical subject area, hence having started with that landscape, and so I am starting with prediction. Here I’m talking here about predicting an individual’s probability to do x or y or z. For example, respond to a marketing offer, or spend a certain amount on a credit card, or buy that pair of shoes. This is familiar to all marketers, both digital and ‘other’, which is nice in that it also serves to make the point that there is little difference between the underlying analytics that drives digital and non-digital marketing. 

The starting point for prediction is the marketer’s brain. Whether it’s selecting customers for a direct marketing campaign, approaching a person in the street, or digitally targeting prospects, all marketers are trying to choose a group of customers more likely to do something. A good marketer can do this with their own insight, without analytics and still obtain good conversion rates. Their brains are predictive algorithms! However, as we are increasingly being told, it is possible to improve upon this.

That next step of sophistication is to analyse the KPI of interest, such as response, and see how it varied across different customer behaviours, such as age, income, time on website etc. Combining these different variables – for example by building a ‘decision tree’ (which creates smaller groups of customers, using more customer behaviours, into higher and lower response groups) – gives more predictive value. We can identify groups of very high responsive customers to communicate with efficiently, and low response customers that we can avoid bothering.

More complex algorithms include regression of some type. For regression, think ‘drawing a straight line through a set of points on a graph so that the line fits best’. I see this as a more sophisticated technique than a decision tree and will often give more predictive value.

That said, there are now new(er) algorithms that improve the accuracy further. Many of the newer algorithms improve upon the straight line, by essentially drawing a curvy line through the data points. This is achieved by building additional models that fine tune the original straight line. ‘Gradient boosting’ is one such class of algorithms that can do this.

Now for clarity, the analysis in these examples is of what happened in the past. The assumption is that customer behaviour will be similar in the future, and only in that sense is it ‘predictive’, there is no magic here, and in that sense, consumers are often scared by AI for the wrong reasons.

As a practitioner of these techniques, i knew that there was much to be gained by balancing the focus on the KPI of interest, as much as the algorithm used. For example, identifying churners that can become high value customers, not just ‘all churners’ will pay more dividends.  Equally, being able to deploy a model within 4 days rather than 4 months brings speed to market and greater bottom line vale again. Therefore, sophisticated automation techniques (across the process of acquiring data, building monitoring and updating models, and deployment of those) are also an important consideration to driving business performance.

The important take away from this is that what we as marketers (digital / non digital / omni- channel) have always been trying to predict things. Being focused on the true business goal, as well as the automation of processes, are as important as which algorithm is being used.

Thanks for the likes on the previous post, (and in advance for any on this). In the next few weeks I’ll make an effort to give a similar ‘degrees of sophistication’ view on one of the subject areas within ‘driving portfolio decisions’ which may highlight some interesting contrasts.