Inherit a model to build a new model I have the following problem.

*

*I have a model, model A. It classifies a persons ability to buy a product. It's a Logistic Regression model.


*I want to add a new feature, but if I add it to model A that feature only have one unique value for every instances. Not much to train on. The feature is an exogenous feature, capturing the state of the economy, one value per year. The other features are information about the persons in the population, like income, employment etc.


*Also, model A, is in production, so I don't want to modify it. I want to build an add-on that builds on model A and include the new feature.


*My thought was to train a new model, model B, on a separate dataset with these state of the economy features spanning several years to get a valid coefficient for this feature (or features) in a logistic regression model. And then add model B to model A in some form of stacking set up.


*My idea is to "inherit" model A, using the returned probabilities, and then stack them with model B.
But I'm unsure about this approach. Given my problem, what would be an expert opinion on how to proceed?
 A: Approach 1: Adding the feature to the existing model

*

*You mention that you wouldn't want to do this because the feature is exogenous and so will have the same value for everyone at a given time. This isn't really a problem so long as your training dataset does have a range of values for this feature and there is a relationship with the target.

*It might also be worth exploring non-linear models as the impact of the economic feature might vary with a persons individual attributes (higher income people less effected by state of economy etc.).

Approach 2: Stacking models

*

*The stacking approach could combine two models A and B. Here you already have an individual model, model A. Then you can train a new model which takes as inputs: the output of model A plus economic features.

Approach 3: Ensembling models

*

*Here, we train a separate model B using economic features and then blend the output with that of model A in some way.

*I don't think this approach is sensible since as you said, the economic features are constant over a given time period.

There's no one right way to do this but my first preference would be to use Approach 1 and incorporate all the features in 1 non-linear model. However, this may not be practical for business/time/explainability reasons. In that case, I would go the Approach 2 route and use the output of A as the input to a second model.
