When facing a wide enough range of explanatory variables, like we are now with our customer data, a reasonable question that should probably pop up is: Which is the subset of variables that maximizes the model's performance? Stepwise regression tries to answer that question. 

It consists of a set of incremental procedures, from which the step part of the name comes, where a different combination of variables are tried out to find out the most satisfying combination. You can think about every kind of stepwise regression as being composed of the following steps:

1. Estimate all possible linear models with a number of n-m variables.
2. Evaluate the best among that set of linear models.
3. Evaluate the overall best performing model among the ones selected within each step.

While the last step will be performed once during the entire procedure, the first two will be performed a number of times equal to the total number of available explanatory variables. To better understand this, we have to talk about two different families of stepwise regression:

• Backward stepwise regression, where the starting point is a model including all the n explanatory variables and during each step the number of employed variables is reduced by one
• Forward stepwise regression, where the final subset of variables is defined starting from no variable and adding one variable from the n during each iteration

I am going to give you some more details in a second, but first you should be aware of the existence of one more kind of stepwise regression that is called alternatively best subset selection, best subset regression, or all possible subsets regression. This technique involves the actual estimation of all possible models based on the set of variables and the final evaluation of the best fitting model. The main downside of this technique is the exponential growth of the number of models to be fitted corresponding to the increase of the number of variables.