I am building a model to predict, say house prices. Within my data I have sales and rentals. The Y variable is the price of either the sales or rentals. I also have a number of X variables to predict Y, such as number of bedrooms, bathrooms, meters squared etc.

I believe that the model will firstly make a split on the variable "sales" vs "rentals" as this would reduce the loss function - RMSE - the most.

Do you think it is best to train 2 models one for "sales" and the other for "rentals"? The RMSE for the model is quite high and this is in part due to the incorrect "Sales" predictions.


1 Answer 1


This is the main advantage of ML: if the variable has any predictive value (that is not included in another variable), it should be used by the model. So, generally speaking it doesn't really make sense to handpick your variables to make different versions of your model (however it make sense to handpick some you want to throw). That would just be equivalent to picking the first binary split in your first tree... you are not achieving much with that.

Edit: ok, it seems that the target is not really well defined as you aggregate things that are monthly payments and things that are the house value. In that case it make sense to have two models. (Honestly it would even make more sense not to aggregate those two distinct dataset in the first place).

  • $\begingroup$ What if the "rentals" variable are around \$1000 and the "sales" variables are \$300,000 combining the two output types give me different errors, an error on the "rentals" might be of magnitude \$100 but the error on the "sales" might be \$5,000 dollars and combining them (I expect) gives me "incorrect" RMSE results. $\endgroup$
    – user113156
    Mar 11, 2020 at 17:12
  • $\begingroup$ On RMSE getting skewed toward the sales prices, just produce two metrics: your test set can be split into two and you can report the two metrics. // While the first split of the first (several) tree(s) is almost sure to be rental/sale, the first tree(s) will have accounted for the large average difference, and later trees might adjust those base prices for both types using only other features. You can help that by modeling log-price instead, to mitigate the heteroskedasticity. Or, likely better, if you can, model house price, and for rentals post-process to estimate rent from house value. $\endgroup$
    – Ben Reiniger
    Feb 9 at 15:26

Your Answer

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

Not the answer you're looking for? Browse other questions tagged or ask your own question.