Neural net to predict stock price (value or direction), I have some issues about cost value that is not changing sensible

Hi there, all mentors and classmates.
I’m recently working on building a neural network to predict price change direction (2 class classification ) based on features like open , high , low and close prices on 5 min time frame and some other indicators , there’s an issue about that the cost value at starting point is so high that is kind of normal but after some iteration (say 25 times) the cost value get close to 0.68 and after that it won’t change sensually , maybe 0.01 after 2000 times, and the accuracy will be some thing like 60% on train data and 70% on test set!
My question is whats wrong with that changing the cost fast in the first iterations then stopping , is the 0.68 the final possible value for my problem?
Thanks
if anyone is interested in working on that field feel free to say me.
( BETTER TO SAY THAT I CHECKED GRADIENT CHECKING AND MY IMPLEMENTATION IS RIGHT )

Learning rate / scale of data might be some issues to look at.

Have you seen this link?

HI there, how did your project go? I am trying to do the same sort of thing.

I was going linear regression before, but thought I would see how this machine learning model would fair.

Allan