Predicting Marketing Engagement - Part 1

The family had our first bout of COVID over the last week. We are all back in fine form now.

I'll start part 1 by doing the following: Cleaning data + finding engagement analytics + segmenting by CLV and months since inception.

The data: The dataset is a famous machine learning dataset for marketing car insurance, you can find it yourself with a simple search on DuckDuckGo for 'IBM Watson Marketing Insurance Data'.

Cleaning the Data

When you are given a dataset, the immediate goal is to discover how clean the data is, there's loads of intellectual capital absolutely wasted on data cleaning. But it's necessary if you're crap at data governance.

From there I do some light analysis to start finding a primary variable and explanatory variables then I see if it's a linear regression problem or a classification problem (level 10 but I have to say it).

Finding Engagement Analytics

The dataset has a column named 'Response' and I'll use that as the primary variable. It's binary, meaning, 2 outcomes (yes and no) are present. So this a good area to start.

We can then group that same engagement response by renewal offers. It's great to see our offers here and what people engage with.

Then I break down to car types grouped by offer because then I start seeing some real behaviour come into play.

Then I go into the sales channels (by car size) and obviously the sales folk are killing it because they are talking to warmed up buyers.

...Also interesting that the mid-size car is most engaged. The must push for volume instead of most value.

One more thing I've done is check 'engagement rate since policy inception'. And it's just good to do a vibe check on the customer experience. There's some good peaks and troughs here.

That 36 month mark has room for some cross-sells. How about pet or home and contents? Then I'd do vibe checks at 40-month marks.

So I analyse types of offers, customers, and how they engage at the bottom of the funnel. I do this to get a total marketing and sales snapshot, just not who is engaged.

Going a bit overboard in early analysis will give a more rounded view in handing off insight to the client once you have a working predictive model.

Segmentation by CLV (Customer Lifetime Value)

Insurance is a funny one, I usually start at gender or taxonomy to do segmentation, I'll start here by policy age.

I'll group by CLV and policy age. Clear significance here when I look at log scale.

Then I'll do a bit more analysis on the segments engagement rate by high and low CLV.

We will always get price buyers, but this looks great to me. So brand is doing it's job very well to keep quality customers.

I'll end part 1 there with defining what is valuable in the data. Next I'll start getting the data ready to model predictions.

Any questions or ideas mash that reply button.