[work in progress 2017-04-15]
A Critical Issue
If we assume that significant seismic events are indeed preceded by distinctive radio emissions generated by geological stresses, and that these patterns may be detected reasonably reliably, there remains at least one significant issue in applying them to the prediction problem. This the time delay between the radio event and the seismic event.
Reports in the literature describe a latencies anywhere between hours and weeks, and it seems likely that this may confound the association between earthquakes and precursors. There could be any number of unknown variables in the physical environment affecting the generation and propagation of the radio signal. For all practical purposes these are effectively random, and so probably best considered another form of noise – in this case in the temporal dimension.
It may be that the radio signal patterns of precursors have recognisable correlation with not only the ensuing seismic events but the length of time before those events. Another (optimistic) possibility is that the neural networks applied will learn this correlation.
At this stage, the best bet seems experimentation…