All the r-value in a correlational study tells you is the strength of the correlation. An r-value of .37 is NOT all that strong. The r-squared value of .137 tells you that 13.7% of the variance in the y-statistic (response variable) can be explained by the value of the x-variable. Basically, the relationship is complete crap. I think Shack has an idea what was going on there...maybe I missed something, but just using my stats knowledge (and I teach it).
A .37 r-value isn't all that strong although I wouldn't call it complete crap. Meanwhile, the discussion is about whether an r-value of .85 is worthless or not.
The assertion that an R-value of 0.1 is useful in predicting the outcome of a baseball game leaves me absolutely speechless. (I spent a minute squinting to convince myself that you really did put a decimal point before the 1... yep, you really did.)
The best gamblers are only right 55% of the time. A model that can give you a small edge is huge.
Here. Suppose we're playing a game. I pick a person from anywhere around the world and you have to tell me whether it's a male or female. And suppose we play this game 100,000 times for a $1 each.
If you have access to google then you'll presume that every single person is male and therefore will be right roughly 50.4% of the time. That means on average you'll be right 866 times more than the times that you'll be wrong and you'll get a nice sum.
Now suppose that I also tell you the country the person is from. In that case, you can use country demographics to figure out when you should pick males (China, India) and when you should pick females (United States, Europe).
Such a method won't have much predictive power. But it's enough that you'll be right considerably more often and mean you'll pocket considerably more money. You don't necessarily need to explain everything for a model to have value.