Answer: a winning lottery ticket.
Just think about it for a moment. The returns are fantastic. There's almost no risk. The minimum investment is trivial. Your money is only tied up for a few days. You get your picture taken holding a big check. What's not to love?
Now, I can hear some of you negative types out there complaining about the difficulty of distinguishing between winning tickets and losing tickets (which are notoriously bad investments), but people routinely ignore these concerns when looking at the business plans behind IPOs. Why should you have a lower investing standard when dealing with your broker than you have at the 7 Eleven?
Take promises of targeted marketing. While it's true that almost all marketing is targeted to some extent and a few companies have been able to take that targeting to a relatively high level,* identifying customers who have a high likelihood (rather than a slightly higher likelihood) of buying something remains an extraordinarily challenging business problem. Most of the proposals you'll read that rely on solving that problem fall into the winning lottery ticket category.
Even with the recent explosion of consumer level data, the vast majority of plans to use targeted marketing run into one or both of the following problems: the lift provided by the selection method won't be large enough; the list produced won't be large enough.
Let's take the Groupon example. As pointed out here by Kaiser Fung, the merchants want new customers who are likely to become regulars. How would you go about targeting this segment? You might try matching the offer with the demographics of website the customers came in through, for instance, high end restaurants for people who came to Groupon through a New Yorker ad, but you'll still get lots responders who are already regular customers and more than a few bargain hunters (yes, even from the New Yorker). Or you could make offers only to people who have been identified as new to the area and are on the mailing lists of similar businesses in their old town, but I can tell you from experience, the number of names you'll get probably won't be large enough to bother with.
And Groupon has to thread an extraordinarily fine needle here. In most business situations, we might have a few customers who end up costing us a little money (for example, someone who just gets the loss leader at the drive-thru), but we're probably talking about fairly trivial amounts. In these cases, it's usually enough to build a model that distinguishes between responders and non-responder and fortunately, response is generally a quick and easy variable to measure reliably.
Groupon has to worry about non-responders (who are still associated with some costs), and about bargain hunters who use the offer then never come back (who cost Groupon's partners a substantial amount), and about regulars who use an offer for a visit they would have made anyway (who represent a double loss).
Separating all this chaff from the customers you want would be daunting even with the best of data and you will not have good data. How do I know? Because I've been there. I've dealt with third party data and I've written the hundreds of lines of SAS code needed to produce clean, usable data-sets. And I was only dealing with data from four or five sources, not trying to tease out a badly defined target variable from data collected from thousands of merchants. (remember, we're trying to identify customers who made their first visit using a Groupon offer and have since returned on their own dime.)
On top of all this, we're talking about a targeted marketing strategy that would have to work with everything from family pizza parlors to upscale wine bars, from pricey spas to summer camps, from teeth whitening to Scotchguarding (all of which have been recently offered by the company).
It's possible that Groupon will get around these problems but, until then, you might be better off with a scratcher-based portfolio.
* Of course, some people have proven pretty good at picking lottery tickets too.