I have to implement a time-series tool in my company, and I'm not sure what I'm dealing with, when looking at the old tool. We have one target variable (sales) and a lot of different independent variables (e.g. spendings or amount of new users).
My coworkers are used to work with an automated blackbox-tool. When they are done 'modeling' they get something like a block of equation terms (screenshot). That consists of a curve fitting transformation to the equation term and the lag of the current series or independent variable. The goal of the modeling is forecasting sales with a holdout set.
The tool itself works with a genetic algorithm, which allows the modeler to steer population and mutation rate only.
I have two questions, and I hope you can help me:
- I know how the concept of lags works, and I know to do curve fitting in R. However, what I really don't know is, what I'm seeing. Are these independent variables just an 'illusion' and we are dealing with a vector automated regression? Or is it just simpler or more complicated, are these independent variables the x from the VARX?
- The second question is a general one about hold-out sets. We often have several years of data which we should model n a weekly base, e.g. three years of data result in 156 data points in the tool.
However, we usually have campaigns which are only active 4-6 weeks in one year and not in the other two years, thus we have variables that are only active in a certain time period. How do we deal with that?
My coworkers still integrate them, although this result in a bad fit for the test data.
How is it possible to integrate a variable that is only active such a short weeks in such a long time series, when we have to do work with a hold-out set? Is there an alternative to the holdout approach?