@Xiang Li: I do have a logic question regarding your excellent post:
As I understand it you feed the NN with same data (technical indictors) you got from your algo.All technical indicators, at least regarding a certain stock/index/etc., are derived from a certain set of timeseries data like stock price (start, end, closing) and volume.
Wouldn't it be much more interesting (and rewarding) to feed the NN the raw data plus the manually selected points of entry into a trade for the same day (buy, sell, short sell, close position, price, perhaps stop loss) and do this for a certain period of time and then let the NN predict it for the test period.
Perhaps one can feed some more external raw information into the NN that may affect the stock price like other stock market indices, macro indicators and so?