The title is correct. That sentence just means that they're predictors, not necessarily that they're positive predictors. As in, if you know one you can predict the other.
In the results section:
> There was a negative correlation between EQ-C and the Calculation Skills composite score with greater empathizing predicting lower math achievement (r(110) = −0.22, p = 0.02).
Considering that the sample questionnaires have to be administered by hand to each child, how would you scale the sample size for a study in this space?
It looks like the study that introduced the EQ-C and SQ-C in 2009 was only run on 1256 "parents of typically developing children" and 265 "children with Autism".[0]
Correlation of 0.2 is negligible with any sample size. The potential for invalid inferences is only pronounced by a small sample size. A samples size of a thousand is much better. But with n=100, r=0.2 there is a bigger chance that local factors influence the result than that the observed association says anything about human nature.
In the results section:
> There was a negative correlation between EQ-C and the Calculation Skills composite score with greater empathizing predicting lower math achievement (r(110) = −0.22, p = 0.02).