This is a good example of what not to do as a data scientist: draw confident conclusions when there is not strong evidence to support that in the data. First o

This is a good example of what not to do as a data scientist: draw confident conclusions when there is not strong evidence to support that in the data. First of all, an R-squared of 0.15 is nothing to write home about. That is telling you that a massive amount of the variance in the data is not explained by whatever field(s) you are looking at. But hey, 0.15 is something! That means it has "some predictive power"! Maybe, maybe not. First of all, saying you've founding something with "some predictive power" is a cop out. Your human editors and writers have a lot of "predictive power" themselves just based on their experience. The real question is if the predictive power of this CTR test is a useful compliment to their judgment, and/or if it is worth all the time/effort you put into it. So let's look at the data little more deeply. A good little trick of analyzing data is to throw out the outliers and then see if that changes your conclusions. This needs to be done with care (sometimes outliers are important!), but when analyzing data you want to play with it a little bit to get a feel for it. In this dataset in particular, what if I throw out the 2 highest pageview stories and the 2 lowest pageview stories? Well the whole conclusion collapses, because the 2 highest pageview stories are basically the only 2 data points supporting this conclusion in the entire dataset. Well crap, that isn't good. At this point I'd stop being confident at all because if taking away a couple outliers destroys your conclusion, that is a bad sign. That doesn't mean the CTR tool is worthless, it just means you can't make the confident claim about it's usefulness from the data. But let's keep going anyways, the idea seems interesting. Here is how to test it. Give access to this tool to half of your writers and not the other half. Let them each write a bunch of stories and compare the page-views of the two sets. Since you can't make this test completely randomized, you are going to have to control for other factors which might be important (a writer's skill level, past experience, etc), but it should be do-able. Get a bunch of data points, plug them into a model and see if the tool had any impact. At that point you should be ready to write a blog post about it.