Here on this paper it is concluded that the periods with large Hurst exponent can be predicted more accurately than those with H values close to random series. This suggests that stock markets are not totally random in all periods. Some periods have strong trend structure and this structure can be learnt by neural networks to benefit forecasting.
I just would like to make the parallel between this study and the study about the low entropy state.
Comments
From fractical perspecitve you can refer to the spinal implant EA that is trying to incorporate the concepts of predictability together with market direction.
http://beathespread.com/file/view/8649/spinal-implant-ea-fractal-scalper-ea-with-neural-net-with-rational-activation-function
http://beathespread.com/file/view/7561/fractal-scalper
http://beathespread.com/file/view/8954/spinal-implant-sigm-ea-with-sigmoidal-transfer-function
Yep, this is the bonus from Beathespread !