Scaling is absolutely essential in order to get rid of spurious associations between dependent and independent variables. The famous example is regressing advertising on sales cross sectionally. You get a huge R-squared and significant slope coefficients, but only because bigger firms have both higher advertising budgets and higher sales. In the limit, if all firms budget advertising as the same percentage of sales, you get R-squared equal to one and a coefficient equal to the budgeted percentage; but it means nothing, other than confirming that bigger firms are bigger than smaller firms: the association is spurious because both the dependent and independent variables are correlated with firm size.

The best discussion I've seen for accounting researchers is Easton's paper (attached). He argues, correctly in my view, that the deflator is ideally driven by theory. Happy reading. I have used this in my doctoral seminar for more than a decade and it never seems to get tired.

you should refer to section 1.4 of  "Introductory Econometrics for Finance" by Chris Brooks. It has discussed the possible ways of rescaling or modelling returns.

In order to reduce high numbers and to convert equally them in low figures, you should use natural logarithm (Ln). 

The rescaling process consists in transforming the original raw data into a normalized set of data with which you can operate and agregate information without creating a distorition in the final results.

Rescaling can be very simple. Some of the actual proceedings are:

1. zero-mean normalization:

By the ecuation: yi=(xi- mean xi) /standard deviation xi

where yi is the normalized vector, xi is the original vector of variables

2. Min-max normalization:

By the ecuation: yi=(xi-min xi) /(max xi-min xi)

And some others alternatives you can find on the internet. According to your explanation I strongly recommend you rescaling given by formula two. With this one, you can manipulate data with results in the interval [0;1], so you can establish easily a new scale for evaluation. Just depends on your interests. There are some more complexs ways to do that, but you will obtain similar results in a different scale or you have to examine new nature of variables since classic math to artificial intelligence-based math. I hope to have been useful to your research

May use GDP to deflate your dependent too.

Go to the R project and select the package dummies

Usually we take the logarithm of large number. This is so-called "Thumb Rule"

I have got the following confusing situation - when I scale the response variable based on another categorical variable ( yi=(xi- mean xi) /standard deviation xi per category/level), there is a better fit to the log-normal distribution. When I leave the response variable as is, non of the error-distributions I've tried fits. But is it , principally, OK to scale your response variable in such manner? thank you!