1
$\begingroup$

I am trying to determine a good tool that will help me generate a probability of a sale for a list of of 300,000 products. I have a table of historical sales data (with about 300,000 records) that contains around 10 continuous variables along with a dependent variable that has a yes/no (i.e., binary outcome) value indicating whether product in the list has had a sale in the past 12 months.

The historical data essentially looks like this.

Product1,2,3 etc
Variable 1
Variable 2 
Variable 3
Variable 4
Variable 5
Variable 6
Variable 7
Variable 8
Variable 9
Variable 10
Sold in past 12 months (Yes or No)

The last variable in the list is of course the dependent variable.

All I want to do is to find a tool that is going to be the best or easiest to use, so that I can assign a probability to each product in the list, essentially giving me the chance to condense my list to the products that are the highest likelihood to generate a sale, so that I can list those products instead of the others that have lower probability of generating a sale.

Ideally, the tool could do a quick logistic regression, or some other probability calculation based on the available variables, and thereby give me a (RVU-like) number (perhaps a probability ranging from 0 to 1) for each product, allowing me to quickly select the top 50,000 products to list on a website, since they have the higher probability of generating a sale according to the available variables.

I am of course assuming that the variables are somehow correlated to the outcome, but perhaps the tool will help me determine that.

  1. Does anyone have any suggestions of a good tool to accomplish this? I would presume that there is a simple way to set this up in Microsoft Excel, but if not, then a piece of software that does this would of course be great too.

  2. I am also open to suggestions as to which type of regression analysis (or other analysis) is the best to accomplish this.

Thanks for any suggestions.

$\endgroup$
  • $\begingroup$ A data scientist could easily do this; typically in R or python. Even Excel seems to be a possibility. My main concern is that your model might be suboptimal; do you really want to know what to know which items have the potential to be sold, or do you want to know how many of each you are likely to sell? The first is a classification problem; the second is a regression. If your items are from the "long tail" and rarely bought, you might be right, otherwise consider regression. $\endgroup$ – Emre Mar 6 '16 at 19:58
1
$\begingroup$

Logistic regression would be an ideal candidate for assessing the probability of a sale, but it would be wise to take Emre's comments on 'how many' into account. In this case even a basic aggregation of your top 50000 with a 'yes' dependent variable would get you there-- but why stop there?

You may want to isolate whether or not seasonal (or other cyclical) factors are influencing purchasing behavior (e.g. if you were attempting to optimize inventory levels for the summer purchasing habits). In this case, linear regression with dummy variables accounting for the season or month may be a good way to delve deeper (this touches on the 'classification' point made by Emre). The more historical data you have here, the better.

Back to your point on 'fast'-- you mention Excel so I am assuming Python or R are being ruled out.

RapidMiner or Knime would fit your requirement for 'fast' in any of these modeling contexts. Otherwise, if you're curious, scikit-learn (python) has about 3 dozen other general linear models freely available (R has a vast library as well).

Attached is a link on how to do this kind of modeling within RapidMiner--

Logistic Regression Example (RapidMiner)

https://www.youtube.com/watch?v=0npgGPK0zR8

Happy modeling!

$\endgroup$
0
$\begingroup$

If you can trancode the data into a regular CSV file, then Orange can open it. Orange is a visual programming tool for data science. It includes several classication and regression algorithms which you can use to easily predict any unlabeled data you might have. There are also widgets for clustering and cluster analysis, which you may find useful (e.g. find if there is a group of products that better fits together than rest).

Orange clustering example

$\endgroup$

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.