I also did something similar, running various algorithms against price and volume history and found the same thing; you can make the algorithm be profitable on old data but it won't be profitable in the real world when applied on new data.

I even wrote algorithms with machine learning and trained them on old data (training set), then tried them on a different set of old data (test set) and they worked quite well, but again, when applied to the real world with the new prices arriving in real time, the prediction was not profitable.

I suspect that this might be because many people are running bots which learn from past data and adjust their strategies; so as soon as enough bots discover a pattern from old data, that pattern stops working.

I got the feeling that it's basically impossible to make a good trading bot just from looking at past price and volume data. I haven't tried to bring news feeds and sentiment analysis into my algorithms yet. Maybe that would help.