Hahaha yup, that's pretty much the definition of "results to send you back to the drawing board".
Reminds me of when someone on SoSH found a statistical model that ranked Trot Nixon as one of the best outfielders in the majors. The overwhelming response was that it must be a pretty terrible model. Sometimes, I think people fail to realize that there is such a thing as bad modeling and good modeling, on both sides of the whole statistics debate. The "statistics are infallible" crowd sweeps it under the rug, while the "numbers lie!' crowd points to the bad ones almost exclusively.
I'm sure a lot of you have built statistical models before, but for those that haven't, there is an explanatory requirement involved. The principle of parsimony dictates that, in addition to increased predictive power, it's also desirable to use as few input variables as you can, creating a tradeoff that you have to balance for. This means that, for each input that you add into the model, you have to give a clear justification, both statistically and logically, for why this variable should be included. This enables you to create simpler and more stable models that have an actual basis in causal relationships between variables, rather than tracing the idiosyncrasies of past data in a way that all but guarantees that they'll have no predictive power against new (future) data.
Look, I could come up with a statistical model that says that JaMarcus Russell is a better quarterback than Tom Brady if I wanted to. It would be a terrible model, because I would have to purposefully defy common sense to build it, but it could be done. But building that model doesn't make its results true, and it doesn't invalidate the awesome work that good statisticians do either.
