Alternative to Q-Q plot to evaluate regression results

[joostneuj]

Alternative to Q-Q plot to evaluate regression results

Task: Extend RegressionEvaluator class to include another graph.

Task Description

Feedback came from data science team at Port of Antwerp. For them, what makes Cobra unique is that most of it's outputs can be analyzed by data scientists but at the same time, can be directly used to start a discussion on business side. But this is not the case for evaluating a regression. Cobra can plot predictions vs actuals and a Q-Q plot. Especially the Q-Q plot is not something you can show business. Some of their remarks:

1. The Q-Q plot (by definition) does not have units which makes interpretation for non-experts more difficult
2. if the 'current model line' deviates from the 'perfect model line', it is still difficult to assess the impact: is this a minor issue business can live with or is it a major concern when we deploy the model?
3. The plot assumes results need to be normally distributed, but this might not always be the case.

So, I suggest to add another plot to the RegressionEvaluator class which should be easier to interpret by a wider audience. I added an example as attachment. In the example I assumed there would be some kind of baseline model, but not yet sure how feasible this is.