I am working with demand patterns and the seasonality of products. For example, I want to categorize products as having a certain selling pattern. For example. Picture a bar graph with months on x axis and unit demand on y axis, forming a bell curve. Are there data mining tools that would tell me x number of products (more or less) have a bell curve shaped demand pattern? Of course, there could be innumerable demand patterns.
Tek-Tips is the largest IT community on the Internet today!
Members share and learn making Tek-Tips Forums the best source of peer-reviewed technical information on the Internet!
-
Congratulations Chriss Miller on being selected by the Tek-Tips community for having the most helpful posts in the forums last week. Way to Go!
Does Data Mining Do This? 3
- Thread starter marcon
- Start date
-
3
- #2
I suggest a clustering algorithm.
Each product is a row in the database. There are 12 fields in each row: the percent of annual units sold in January, February, etc. It is important that you use percent of annual units to form a common scale.
Cluster the rows into groups using the 12 fields as clustering attributes. This will give you clusters of products having similar selling patterns.
Then create graphs of the average curve for each cluster. This will have the effect of smoothing the curve and making it more recognizable.
Examination of those curves will reveal selling patterns. Look at the members of each cluster to tell you what might explain the selling patterns.
Each product is a row in the database. There are 12 fields in each row: the percent of annual units sold in January, February, etc. It is important that you use percent of annual units to form a common scale.
Cluster the rows into groups using the 12 fields as clustering attributes. This will give you clusters of products having similar selling patterns.
Then create graphs of the average curve for each cluster. This will have the effect of smoothing the curve and making it more recognizable.
Examination of those curves will reveal selling patterns. Look at the members of each cluster to tell you what might explain the selling patterns.
Accounting for promotional factors
If you know that promotional factors influence unit sales, then you would take them into account in the design of your analysis. I might build a simple predictive model (such as a multiple regression model) with unit sales as the target variable and promotional features (ad frequency, price reduction, ad size, media circulation, gross rating points, etc.) as predictor variables. Once you build a model having a good fit, you can "remove the influence" of the predictor variables by computing the expected demand with certain promotional variables equal to zero. This would place all products on equal footing, and you might then continue to cluster products.
Another method is to begin with a clustering exercise as discussed above. I mentioned up there you might study which products had similar sales patterns and speculate on reasons why. Merchandising and marketing people might notice that products appearing in the same clusters were all advertised in a fall circular. You'd smack your forehead and say of course their similarity in sales patterns is the result of having similar promotional characteristics. This might or might not be important to you, depending on what your company is doing strategically. For example, you might notice that certain products DO NOT appear in the cluster marked by heavy promotion, even though they SHOULD. In this case, you have discovered products for whom promotions aren't working.
On the other hand, if your strategic goal was to find out more about the consumer's habits - to understand how some products are purchased periodically or in certain seasons - then you might wish to "control" the promotional factors. One way to do this without resorting to statistical regression analysis is to group products into those that are promoted versus those that are not. Or those that receive promotional strategy A versus those that receive strategy B versus C. You might then cluster products within these groups, and you could be sure that the products that show up in the same clusters are not there due to promotional factors (something THE COMPANY does) but rather due to demand factors (something THE BUYER does).
Notice that the act of grouping products into categories based on their promotional features is itself an act of clustering. If you have good promotional data (which most firms do not), then you could get fancy and cluster the products in 2 stages, first by promotional patterns using promotional data as attributes, then by sales patterns within promotional clusters using unit demand distribution as the attributes.
My opinions anyway.
If you know that promotional factors influence unit sales, then you would take them into account in the design of your analysis. I might build a simple predictive model (such as a multiple regression model) with unit sales as the target variable and promotional features (ad frequency, price reduction, ad size, media circulation, gross rating points, etc.) as predictor variables. Once you build a model having a good fit, you can "remove the influence" of the predictor variables by computing the expected demand with certain promotional variables equal to zero. This would place all products on equal footing, and you might then continue to cluster products.
Another method is to begin with a clustering exercise as discussed above. I mentioned up there you might study which products had similar sales patterns and speculate on reasons why. Merchandising and marketing people might notice that products appearing in the same clusters were all advertised in a fall circular. You'd smack your forehead and say of course their similarity in sales patterns is the result of having similar promotional characteristics. This might or might not be important to you, depending on what your company is doing strategically. For example, you might notice that certain products DO NOT appear in the cluster marked by heavy promotion, even though they SHOULD. In this case, you have discovered products for whom promotions aren't working.
On the other hand, if your strategic goal was to find out more about the consumer's habits - to understand how some products are purchased periodically or in certain seasons - then you might wish to "control" the promotional factors. One way to do this without resorting to statistical regression analysis is to group products into those that are promoted versus those that are not. Or those that receive promotional strategy A versus those that receive strategy B versus C. You might then cluster products within these groups, and you could be sure that the products that show up in the same clusters are not there due to promotional factors (something THE COMPANY does) but rather due to demand factors (something THE BUYER does).
Notice that the act of grouping products into categories based on their promotional features is itself an act of clustering. If you have good promotional data (which most firms do not), then you could get fancy and cluster the products in 2 stages, first by promotional patterns using promotional data as attributes, then by sales patterns within promotional clusters using unit demand distribution as the attributes.
My opinions anyway.
An alternative would be to have a no-promotion control and then model the uplift due to the promotion.
This may be straight forward in a purely direct invitation campaign (email, direct mail or outbound telemarketing).
If there is in-store activity it may be possible to regional no-promotion controls. This is more difficult as you have to account for regional bias.
Could you advertise a different price for a random group of potential customers (i.e. they dont have to log-on first) on the internet? How do you ensure that the price is the same if/when they next visit the site?
With different channels the consistency issue becomes a problem.
I'll stop rambling now.
- Thread starter
- #7
Thanks for the feedback! jtm111 has pretty much described the choices as they actually exist. On one hand, I am being urged to piggy back a Six Sigma project onto what I am working on, but realistic modeling the of variables and communicating that complexity to customers and management is an issue. Experience with that tells me that valuable time will be spent repeatedly explaining why things are vague.
And on the other hand, much can be gained from looking at the patterns and making real decisions based on it. This is what I have thought all along.
Does anyone know of particularly good sources for clustering algorithms?
There is some potential AKAK, to use control and I think it will become important downstream when more experience exists.
Thanks Again!
And on the other hand, much can be gained from looking at the patterns and making real decisions based on it. This is what I have thought all along.
Does anyone know of particularly good sources for clustering algorithms?
There is some potential AKAK, to use control and I think it will become important downstream when more experience exists.
Thanks Again!
There are two things that you need to do in order to perform meaningful analysis:
1. Plot sales per month per product
2. Get promotional/marketing program history, again per month per product. Collect all information that you can... what type of a discount was offered, for how long, etc. Label the plot if you can so that you can visually associate sales data variations with specific marketing programs.
...then and only then you can start your analysis.
manolito
1. Plot sales per month per product
2. Get promotional/marketing program history, again per month per product. Collect all information that you can... what type of a discount was offered, for how long, etc. Label the plot if you can so that you can visually associate sales data variations with specific marketing programs.
...then and only then you can start your analysis.
manolito
- Thread starter
- #9
- Status
Similar threads
- Locked
- Question
- Locked
- Question
- Locked
- Question