Over 90% of data science models never make it to production.

This internet fact is misleading - the nature of R&D is that much of the work yields no results and gets abandoned. But the lesson is valid. Too much data science work loses track of the original problem statement.

So I use the same framework on all my data science projects to ensure we stay on track.

Connect your dependent variable to a value equation

And here's a simple example.

Imagine a model for a subscription-based software product that predicts customer churn. Evaluating the accuracy of your model is too simple. Do this instead:

• Segment the population of customers into cohorts

• Calculate annual revenue $ and churn % in each cohort

• Calculate churn reduction $ savings from actions taken due to model predictions