NonLinear Forecast/Prediction

mafie12mafie12 Member Posts: 3 Contributor I
edited November 2018 in Help
Hej guys,

I`m currently working on a pricing-issue where I want to forecast the price-decrease of several products.

THE PROBLEM: The historical data points out that the price decreased non-linear so far. Instead it decreases quite strong at the beginning but the longer the product is on the market the more the price approaches a zero decline. I`m pretty much a beginner with Rapidminer and could only figure out that linear regression would not help me to forecast scoring-datasets in a meaningful manner.

Can you help me with that?



  • Options
    mafie12mafie12 Member Posts: 3 Contributor I
    I would be happy for any small hints to get started because running through several Youtube Videos and the Manual didnt help so far.
    Probably I also explained the problem not got enough.

    The price decreases over time as follows (just an example):

    01.01.2012 - 350€
    01.03.2012 - 300€
    01.06.2012 - 270€
    01.09.2012 - 245€
    01.12.2012 - 227€
    01.03.2013 - 215€
    01.06.2013 - 205€
    01.09.2013 - 197€
    01.12.2013 - 191€

    What I need to do is a prediction for the upcoming months. It seems like a logarithmic trend but I dont know how to build such a process effectively.
    Thanks for any advice or sharing of your thoughts.
  • Options
    MariusHelfMariusHelf RapidMiner Certified Expert, Member Posts: 1,869 Unicorn

    if you know that the descent is logarithmic you can try to transform the data before applying a Linear Regression with the Generate Attributes operator, e.g. by taking exp(price).
    In the general case you can use one of the various non-linear regression algorithms like SVM with radial kernel or Gaussian Process.

    Best regards,
  • Options
    mafie12mafie12 Member Posts: 3 Contributor I
    Hej Marius,

    BIG THANKS for your response. I haven´t expected anyone answering anymore and unfortunately I didnt get any message that I fortunately got a response. Anyways, I tried the SVM as well as Neural Networks and it works good for my purpose. I would like to understand your advice to use the exp(price) function for applying linear regression. It seems that it almost straightens the price function except for the first point (300€) but how do I proceed from their? How should I transform the testing/scoring set to apply the regression and get a reasonable forecast?

    Furthermore I was wondering how I can arrange the algorithm to measure seperate values according to the ID. I have product A, product B and product C in my dataset and each has different prices and accordingly different price curves. I indicate the product name as ID but when I run the SVM it can´t distinguish between the different products and gives just a mutual value for each product as if they were al the same. What am I doing wrong in this case?

Sign In or Register to comment.