Interesting, what about a penalizing coefficient integrated into the error function to better predict direction? predicted direction having a smaller constant.
Agree with what Johanmul has stated. Time of day, product and specific microstructure will all affect your prediction accuracy. Also big difference between 1 and 5 min prediction. But is there any reason to force your prediction to these arbitrary time frames? You may already be doing so, but I'd think for the features you are playing with, you'd find better success predicting next N ticks, where median N<2 and avg holding time in the tens of seconds.
Hi garachen, Why do you consider only a market taker approach? You are likely to earn much more just quoting small tick instruments by transforming your pricing model. If you are interested PM me.