The reason you are getting another unusual point is because as you keep removing points, and since these may be outliers if I'm not wrong they have a large impact on the mean and hence once removed, there is a shift in the mean and you'll keep getting outliers.
R^2 only tells you how much of the correlation is accounted for by your statistical model, hence you are right and it should move nearer to 1.
Remember that R^2 is coefficient of determination and NOT your correlation coefficient.
Now with regards to your research, a good predictor would be a combination of body fat.
I'm not sure how statistical you want to be but, to break this down, you have
body fat % = forearm+abdomen+height+weight+hip. Thats a linear equation.
I would suggest you do a step wise formulation by adding and removing variables to see when you get a significant effect. Remember when it comes to weight and body fat, there is usually a combination of predictors.
Transforming data does help depending on how much variation you have between the measurements.
Does this help or have I deviated from the main topic?!