The need to use our heads is why there are so many companies still struggling
with post-event analysis. It’s so much work! Typically, as the joke goes, “post-
event analysis” consists of the salesperson pulling out of the store parking lot,
looking in the rear-view mirror, seeing the big banner in the store window,
and saying: “Yup, we’re good! I got the display!” — Dale Hagemeyer

Only when we get to where we can do post-event analysis in a meaningful way will we truly be able to evaluate an event. Once we can do that — once we know we’re pulling in all the key pieces and thus can establish some causality — we will actually be able to predict the future.”

ALEX RING: I agree. But I would challenge you to think about these things: In a great many cases, we still don’t have the data to drive the kinds of things you’re talking about –- or at least not to automate them.

Consider diverting. We all know that, depending on what category you’re in, on the case count, on the product weight or the price point, there can be a little or a lot of diverting. And diverting can trick lift models.

For example, if you have no syndicated data to match up with shipments, a diverted product can lead a model to “think”: ”Wow! It all sold through!” Likewise, if you integrate syndicated data with shipment data, it can unfairly predict that you have too much or too little inventory back in the warehouse or distribution system.

So diverting, and changes in diverting behavior, can make this kind of analysis irrational.

Another thing: If you’re in foodservice, what do you do about retail operators? For most companies, the direct customer is a wholesaler. You don’t know where your product is going, ultimately. You don’t have data from those indirect customers. Thus, loading into an optimization model the data necessary to do the calculation is akin to a car with a fuel calculation device –- but it’s not hooked up to the speedometer or gas tank. So all it can tell you is: “Well, you have been in the car about an hour; I think you have used about a gallon of gas.”

Therefore, the information is at best directional -– not absolute. To put it in perspective, you have to know where the data gaps are.

This brings up another point: From a syndicated data standpoint, your company’s definition of a Kroger/Cincinnati KMA may be different than, say, an IRI definition or a Nielsen definition. So while the gas gauge may say, “You have 85 miles left,” in your

head you know that you’re driving up the mountain to a ski resort, and that your gas mileage will not be what it was on the flat area behind you. So you don’t have 85 miles, after all.

Again, it’s a case where human intelligence has to be brought to bear, adjusting for a difference the computer cannot account for.

DALE HAGEMEYER: I must concede there’s another big challenge for predictive modeling: change management.

As we all know, too often people feel: “I don’t want to use any of those stupid tools! I don’t even want to ‘Push Here, Dummy!’ I like doing my promotion plans on cocktail napkins, or while teeing up with the customer on the golf course.”

Even you and I, Alex, are not really looking at the change management issue, which is very complex.

Another complex challenge: Collaboration — the industry “C” word that, for too many players, still means: “Please hold still