It could be random. Or, the initial decline the edge showed, relative to backtests, could have been the random part.
Or, subsequent refinement of the method, using post-trade analysis and further development of tools to accompany it, could reveal periods of degradation which could be avoided...