Without giving away specifics, do you forecast price changes, price direction, horizons, etc? I used to think this was impractical but I've changed my mind. Taking first differences of prices is a very high frequency transform which removes lots of valuable information, as a test I used a moving-detrend by taking differences from a 5-minute moving average and then forecasting those. Here is an OUT-OF-SAMPLE forecast, this forecast was generated from data previously unseen and is unbiased. Not shown here is the increasing error bounds as the horizon increases. Also, no, I didn't just pick a particular starting point that had a good fit, the results are robust from most any starting point. How did I do it? All I can say is that Mandelbrot is right.