Thanks for dropping by, Kyle, and for your comment...

2008-04-18T07:49:00.000Z

Thanks for dropping by, Kyle, and for your comments.

Your absolutely right, of course, that in a retail (or any demand generation) application, it's really the expected incremental value (profit) that you want to model. I have used the approach in retail to tackle this problem, and there are two basic approaches. (I'll add something to the FAQ on this too, some time.) Which approach makes sense depends, to some extent, on what you believe the nature of the impact of your action is likely to be.

It seems to me that there are two underlying possibilities. One is that the action drives frequency, i.e. makes people transact more; the other is that it drives purchase size. Of course, the reality may be a combination of the two, but it helps me to simplify.

If we believe that fundamentally the action drives frequency, then a natural approach is to use two-stage modelling, with a binary uplift model predicting whether someone will purchase during the outcome window, and a value model on the purchases to estimate the size. (Of course, this only works if the purchase frequencies are low enough that not many people will purchase more than once during the outcome window.) I think this may be what you were referring to.

You then get something like

E (incremental sales) = Uplift * E (value | purchase)

where Uplift is the increase in purchase probability resulting from the action.

The alternative approach focuses on the idea that you drive purchase size. This is more appropriate either if that's what you believe the basic effect is or if the frequencies are such that the two-stage approach isn't sensible. In that case, I just try to model the uplift in sales volume directly, i.e. I model

E (value | treatment) - E (value | no treatment).

The simplest approach to this is, of course, just to model the two populations separately, with all the problems associated with that. And then there are direct approaches, which I concentrate on.

There are several possible ways of combining the two approaches, but I've never actually tried that.

The other comment I'd make is that, while I have had some success with uplift modelling in retail, particularly catalogue retail, it definitely seems to be harder to make a big impact here than in telco and finance. This is not so much because the approach doesn't work as because, in my experience, there often seems to be a fairly strong, positive correlation between purchase probability/frequency and incremental purchase probability/frequency in retail. So if the approach is to target (say) the top 3 deciles, even though an uplift approach will often make a much more accurate estimate of the impact, it often seems not to pick very different people. And, of course, it suffers slightly from being a second-order modelling approach, so what it gains from modelling the right thing, it can lose in the noise.

But don't let that put you off. Overall, I've still had better results in retail using uplift approaches than traditional ones.

Came across your blog while googling with the term...

2008-04-18T00:50:00.000Z

Came across your blog while googling with the term "incremental modeling" (meaning modeling the incremental response or sales) and was really glad to find your blog.

I didn't have time to go through all your posts yet but I was wow'd by the first couple of posts. Good stuff, keep it up Nicholas!

I work for a retailer which too is facing the same incremental (or UpLift) challenge as TelCos and Banks. However, as a retailer our world is slightly different. For banks or TelCos, identifying the most influence-able customers solves the problem-- As long as you capture/save a customer through your UpLift Models or whatever modeling technique, you lock in a relatively steady revenue stream from her/him. Everyone goes home happy.

In the retail world, simply finding the most influence-able customers doesn't quite solve the problem, because the more influence-ble customers are not necessarily the better spenders. I may very well end up sending a catalog to someone whose response is highly depending on receiving the catalog but would not spend much with us.

We've been testing two-model incremental modeling (the treatment-model-minus-control model approach. Very backwards, I know) for a year now, and I have seen quite a few campaigns where I have better (incremental) response rate but lower (incremental)sales compare to regular response modeling. It's nice to see the two-model approach still finds good incremental responders, but it's not maximizing what we really want: incremental sales.

I know I can always marry the incremental model with order-value models, but that's going to worsen the problem of noises/interactions between models, so I'm not sure that is going to help.

Anyways, my point is, in retail the real holly grail is this:

Finding customers whoes SPENDING is the most influence-ble by our marketing efforts.

Do you think UpLift Model is a solution that can solve this vale/sales challenge?

Comments on The Scientific Marketer: Uplift Modelling for Cross-Selling: White Paper Av...

Thanks for dropping by, Kyle, and for your comment...

Came across your blog while googling with the term...