Now that you have this new tier change feature, what would it look like to group the data by tier change and then plot the value counts of the commit series? Take a look at this countplot and see what relationships you notice.

This new tier change series that you just engineered is a very informative feature. It looks like when an investor is promoted, there are actually 0 declines in the entire dataset. And when an investor is demoted, they almost always decline. Notice the number of demoted decline relative to none decline. Demoted investors represent two thirds of the declining investors in the dataset. Even though demoted investors are a minority of the total observations. By engineering this new feature you just gave your machine learning algorithm a new tool that will make it much more accurate in predicting investor behavior than it ever would've with the dataset you received originally. That's the power of great feature engineering, but if you want to engineer useful features, you have to use your head to really think about the underlying relationships that you're trying to model and how the features you have might be transformed into something that's even more informative.