Simpson's Paradox for $R^2$

Question

I've been reading about Simpson's Paradox and I understand how if you fit a regression on a per-group basis versus on the entire set, the coefficients in each group can be drastically different from coefficients on the global set because of confounding variables. However, one thought I had was whether such a principle would hold for an accuracy metric. For instance, if I run a linear model on the entire set with an $R^2 >0$, is it possible to split the dataset into two groups $A,B$ such that $R^2_A < 0$ and $R^2_B < 0$?