I had a quick look, here is what I understood:
Their raw data is the price and the first few order book entries at OKCoin, at 10-second intervals, from Feb/2014 to Jul/2014.
To predict the change in price in the next 10 seconds, they take the last 180 to 720 data points (30 to 120 minutes' worth of their data set) and look back in history for occasions when the price wriggled in similar ways. They look at what the price did just after those occasions, and use that information to predict what it will do next. The complicated math is details of how to compare the recent wriggles with the historical wriggles, and combining the "past futures" into the "next future".
They did not actually trade on the exchanges. They used the first 4 months of data to adjust the parameters of their method, and then simulated how it would have performed over the last 2 months (Jun-Jul/2014).
Some comments:
I may have misunderstood, but it seems that they did just ONE test run, using the last 2 months of data.
The performance in that test (doubling the USD capital in 50 days) would be impressive for stock market trading, but for bitcoin trading it seems quite modest, given its hight volatility. If God traded every 10 seconds, as in their simulation, He would probably get that return in a few hours.
In fact, doubling the capital in 50 days may be within the noise level. That is, if several traders traded at random, once a second, for 50 days, some of them may well double their capital, some may lose it all. (I don't know whether this is true, it would have to be checked.)
So, here is one thing that may be wrong with the paper. The method has several parameters. Some of them are meant to be ajusted by "training" the mthod on part of the data, as they did. Others are ostensibly fixed, like the duration of the windows used ("30, 60 and 120 minutes"). But the latter are arbitrary, and they probably tried several values until they got values that seemed to work. But in doing so they may have only selected a "random trader" that was lucky in that test.
They may have discussed this potential flaw in the paper, in the parts that I skipped.
Poll
