Given the state of the economy and the trajectory of the catalog / e-mail marketing industry, I thought I'd give MineThatData Nation access to a presentation I prepared for the CEOs who recently inquired about partnering on a Multichannel Forensics project.
Multichannel Forensics Presentation For CEOs (pdf format ... fewer than 30 slides).
The slides outline what a typical project looks like, given the requests of multichannel CEOs this spring, and roughly outlines project costs!
If you're a member of the e-mail marketing community, there's content at the end of the presentation for you, too!
Thanks Kevin,
ReplyDeletean excellent presentation full of great ideas. I especially liked the description of the cube design. I'm trying to do something similar for our business. I'm also slogging my way through market basket analysis so we can target merchandise preferences.
However I worry that we are so small that we'll not be able to draw valid conclusions from the data. For example, we are currently gathering mailout info to match against purchases. Eyeballing the sales figures I get that sinking feeling - because our organic monthly sales for particular products can be measured with one hand I'm becoming less confident that we can measure lift in any meaningful way.
In your experience can small firms use statistical methods to extract actionable meaning from sparse data?
The size/ throughput of the firms you deal with is truly mind-boggling to me.
Cheers
Oh sure, you can use statistical methods to extract meaning.
ReplyDeleteIn fact, you get to have the most fun when you're working at a small firm!
Certainly, there are conclusions that you cannot draw. It is very hard to do mail and holdout tests, it is even harder to read those results at a merchandise category or department or division level.
Almost everything I describe in the presentation I do for small firms --- I have numerous clients that do $10,000,000 or $15,000,000 sales per year, and I can do just as good a job for them. In fact, I can often do a better job. I can spin through a dataset for a $15,000,000 business in twenty seconds ... it might take two hours to do the same analysis for Nordstrom's 7 million twelve-month buyers.
With sparse data, the analyst plays a much bigger role --- the analyst has to make decisions whether the data is reliable or not --- whereas when I was at Nordstrom, I'm analyzing one group of 500,000 customers against another group of 1,250,000 customers --- there isn't any fun, the data is self-evident, the data makes the decision for you.
Hi Kevin,
ReplyDeleteQuestion about the Email segmentation. When you classify into the one, some and all divisions, are you talking about divisions within, say, Women's Merchandise, or across Women's Merch, Men's Merch, Footwear, etc? And would you also have a forth column for 'no division?'
Another way I'm thinking of it is, if each row is a customer, then could each column be a merch division and the value in each column indicates if there are no purchases, all purchases, or some purchases in that merch division?
I don't have the horsepower to do cross-category purchase propensity, so what would you think of targeting based only on the divisions purchased from, without the predictor of what other divisions might be likely?
Thanks,
Jim
Hi Jim, your ideas are great, don't worry about the logistics of how I set up the segments. If you target based on the merchandise divisions purchased from, you'll maybe get a twenty percent lift in performance ... maybe a 100% lift with some customers, maybe a 0% lift with others, averaging out to a 20% lift.
ReplyDeleteAll of my clients ask me to create the table in the presentation a bit different, based on their unique needs. When they don't give me direction, I let the computer tell me what makes sense, and when I do a factor analysis, I find that type of merchandise (womens vs. mens) is one dimension, buying from one or multiple merchandise divisions (here you might be looking at categories within divisions) is important.
Hi Kevin,
ReplyDeleteI was wondering why you would use a Probit model to calculate the Organic Model scores. Most likely this is due to the fact I am having trouble remembering what Probit is!!
Thanks,
Ted
Probit models tend to work a little bit better when you dealing with probabilities around 0.5 --- whereas logistic regression tends to work well with probabilities around 0.1 or 0.9.
ReplyDeleteAny method that gets us to the desired result (predicting the percentage of sales that are organic) is perfectly fine!