If I have a qubit state $\lvert A\rangle=\frac{1}{\sqrt{2}}(\lvert u\rangle - \lvert d\rangle)$ and I then trying to determine the following expectation value
$$\langle A\lvert \sigma_x\rvert A\rangle$$
I got $<\sigma_x>=\frac{1}{2}(\langle u\rvert - \langle d\rvert)(\lvert d\rangle - \lvert u\rangle)=\frac{1}{2}(-\rangle u | u\rangle - \rangle d | d\rangle)=-1$
But $\lvert A\rangle$ is not really pointing in the $\pm x$ direction, thus how can an expectation value of -1 be obtained (since to get an expectation value of -1, the probability to get +1 must be zero)?