Predictive Modeling for forcasting leads.

  • Posted by a hidden member.
    Log in to view his profile

    Jun 08, 2011 6:34 PM GMT
    Lots of smart guys on here so I thought I'd throw this out there. Does any one have any experience with Regression analysis and predictive modeling? I have some work to do for predicting lead data, unfortunately there are some pretty big outliers in the data that prevent me from using liner regression. Any direction for a learner would be very appreciated icon_smile.gif Danke.
  • Posted by a hidden member.
    Log in to view his profile

    Jun 08, 2011 7:28 PM GMT
    Is it possible to see your data ?
    There is several ways to handle outliers, but it require hypothesis about the source of the outliers and their distribution.

    it's important to know if your outliers are :

    - error in measurement
    - large variability of the measurement process
    - large variability of the observed phenomenon
    - not outlier, ie, you model is not linear

    it's usually a mix
  • ursa_minor

    Posts: 566

    Jun 08, 2011 7:45 PM GMT
    possibly read on Artificial Neural Nets (perhaps you local computer library or google online). I haven't monitored the progress of this technology for the last 10 years but it is a promising field.
  • NerdLifter

    Posts: 1509

    Jun 08, 2011 11:11 PM GMT
    Aye, as Minox said, possible to see the data? It'll also depend on what is the intended application of the data.
  • Posted by a hidden member.
    Log in to view his profile

    Jun 09, 2011 12:31 AM GMT
    Sample data is very small, but here it is:



    Affiliate Lead Report

    Affiliates Cost Per Lead Rev per lead Jan-09 Feb-09 Mar-09 Apr-09 May-09 Jun-09 Jul-09 Aug-09 Sep-09 Oct-09 Nov-09 Dec-09
    Affiliate 1 $5.00 $8.00 387 327 471 399 251 210 426 365 501 417 600 513
    Affiliate 2 $6.00 $8.00 84 74 24 20 4 84 74 24 20 4 84 74
    Affiliate 3 $3.00 $6.00 97 86 136 114 88 83 109 96 77 67 12 12
    Affiliate 4 $4.00 $4.50 233 200 133 112 0 0 0 0 30 24 45 34
    Affiliate 5 $5.00 $9.00 70 52 88 60 57 43 58 48 33 25 0 0
    Affiliate 6 $6.00 $10.00 41 38 55 54 39 34 34 31 22 19 0 0
    Affiliate 7 $7.00 $11.00 4 4 2 2 0 0 2 2 1 1 1 1
    Affiliate 8 $4.00 $7.00 197 152 210 148 150 120 165 124 211 159 175 138
    Affiliate 9 $10.00 $13.00 15 14 11 20 16 18 15 14 11 20 16 18
    Affiliate 10 $12.00 $9.00 35 28 45 41 31 23 37 31 54 49 97 85

  • Posted by a hidden member.
    Log in to view his profile

    Jun 09, 2011 12:33 AM GMT
    Here's where it gets fishy, the predictive data is done with two conditions, 1 with only the odd set of months data, and the second with only the even months data.

    R^2 only gets to 0.4884 and 0.6886 respectively.
  • Posted by a hidden member.
    Log in to view his profile

    Jun 09, 2011 12:39 AM GMT
    I do see some bears on the road in general with predicting leads. Because no two programs are the same, do you really think you can find the values (constants/variables) to build a working model that applies across programs?

    Back when I worked for this yellow pages joint I built SEM click forecast models... it was complete and utter crap, I just built them cause some people wanted me to. I did the only sensible thing when you're dealing with a marketplace like that, which is speculating based on historic data - in which you basically out-model any quality improvements you might make yourself icon_surprised.gif
  • Posted by a hidden member.
    Log in to view his profile

    Jun 10, 2011 3:35 AM GMT
    Not sure you can foercast anything good out fo your datas.

    You have three incomplete record, for affiliates 4 5 and 6, (set of 0s)

    You have affiliate 7 with values to low to distinguish a trend from variability.

    Affiliate 1, 8 and 10 (and 3 up to august) show similiar 'up and down' variation from month to the next, while having different trends.

    If this variations result from seasonal demand variation, then you can 'extract' that seasonality, remove it from the data, then extrapolate and re-apply the seasonal factors on the result.

    But you would need at least two years history to know if it's seasonal effect or not.

    If it's not seasonal effect, but rather identical reaction to demand change (promotions, competitors etc..), then you could still (barely) calculate and remove that effect to calculate a trend, but won't be able to predict future monthly value as you won't know how the demand change effect will be in the future, and it account for variations about 50% of the total.

    If you normalize your data (divide each affiliate value by the serie max) and plot them, you will see there is no clear trend.

    If you think there is some seasonality, and/or can put the hand on more years, I can tell you how to extract and use it.

  • havingfunmtl9...

    Posts: 258

    Jun 10, 2011 3:51 AM GMT
    Use the Tabachnik and Fidel method ! All depends on how you intend to use the data, but you can either eliminate the outliers or run some additional analyses to establish the extent to which they will influence the overall shape of your data.