0
$\begingroup$

Sorry for the bad title, I can't find a good one. So I will try to explain what I'm looking for.

I'm doing sales forecasting with a Regression Forest. (Spark - Scala for the technology) I've worked on some test data and I did my forecast using training data. But some of the features which I have used can't be employed to forecast the future as they would not be known to me at any given time. For example the numbers of customers of a day, their categories, what kind of advantage they have etc.

Do I have to find others features that will be as useful as these ones or Do I need to perform prediction on these features before my sales forecasting and use the predictions? Are there any another solutions? Also, what kind of algorithms should I use for the "features forecasting"?

$\endgroup$
1
$\begingroup$

The question is poorly phrased, I've tried to edit it to the best of my abilities. However here are the problems you've stated,

  1. Some of the features which are being used now can't be used later because they might not be known. Will it affect the model?

  2. If they can be used, what type of algorithm can be chosen?

The answer to the first problem, you have to check the accuracy first before making choosing any new features if the rest of your features give a good enough accuracy then there is no need to choose new features.

The second problem, to predict those values of the features you are using now if they are discrete in nature try classification algorithms whichever fits the model best, else try something along the lines of regression if the input values are continuous. And then use these predicted values along with your existing model and check how the accuracy varies.

$\endgroup$
  • $\begingroup$ Thanks for the edit and your answer. The accuracy is much better with these features that can't be used in the future. These features are about the customers like what kind of customers and how many bought this specific produit. I don't really know if I can use a classification algorithms in this problem. I will try it and see how I can build it. $\endgroup$ – KIToRe Sep 4 '17 at 7:01
  • $\begingroup$ If you can't get values to those features or use them later on, then it's meaningless to consider them in the model. Even if the accuracy is high now, if you can't use them later it doesn't make sense to consider it. According to how you described these features, you can't directly predict them unless you have a statistical model, so either find a way to get those values before hand about the customer or go about in another way. $\endgroup$ – Rahul Aedula Sep 6 '17 at 16:08

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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