As I remember, it is considered stylized fact in academic that returns are positively autocorrelated in short time frames (I believe this means <= 10 min) while it is negatively autocorrelated in longer time frames (>= 30 min), and this does not limit to stock market.
Actually they are negative in short time frames b/c of bid ask bounce. It's a simple mathematical proof to show that even with purely random price series you will observe negative autocorrelation as long as there is a positive bid-ask spread. I believe the original proof was by Jeniffer Conrad at UNC. Outside of intraday periods, at the slightly longer intervals (days to weeks) there are reversals, at monthly intervals there is momentum and at 3-5 year intervals you get negative autocorrelation again.