1
$\begingroup$

I have a large dataset of people, with 30 categorical and continuous characteristics (e.g. age, gender, etc). They also have a binary status given to them each year, for example saying if they bought a product that year or not. The dataset is across 10 years, and the status variable is the only one that changes between years.

I use the characteristics data, plus data on a person's status in previous years, to predict a person's status in future years. The model is based on clustering and makes predictions based on people with the most similar characteristics.

I have now been given a time-series of macro-economic variables (e.g. GDP) that covers the same 10 year time period as the data mentioned above. There is also forecasted GDP for future years, and I have been asked to apply those forecasts to my model. i.e. A lower GDP might mean fewer people will buy the product, but this is more likely to affect people with certain characteristics than others.

I can't think how to incorporate this into my model. The GDP is the same for everyone, so how can I work out how changes in that variable affect a person's status probability?

My concern is that although the feature of interest (GDP) differs between years, it is identical for the millions of instances within each year, and so will be a poor predictor in any model.

$\endgroup$
3
  • $\begingroup$ if you are using linear model take interaction between GDP and binary variables as a new feature. If you are using tree-based models, it is ok to throw GPD with binary features to model. If you are using tree-based models make sure your time-series features are stationary. $\endgroup$ Mar 3, 2021 at 14:48
  • $\begingroup$ How can you have an interaction between a predictor and the response (in LM)? I thought interactions were just between different predictors. $\endgroup$
    – rw2
    Mar 4, 2021 at 11:10
  • $\begingroup$ I was talking about interactions between "characteristics data, plus data on a person's status in previous years" and macro variables. I was not referring to interaction between target and features $\endgroup$ Mar 5, 2021 at 12:05

0

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.