An algorithmic framework for investigating the temporal relationship of magnetic field pulses and earthquakes applied to California

An end-to-end algorithm is described wherein field-collected magnetometer time series data were processed and analyzed for potential statistical correlation with pre-seismic activity. The process included windowing the data, extraction of statistically-determined anomalies via a short term average - long term average (STA-LTA) signal processing technique, collating and ranking the anomalous windows as precursory behavior, and testing the results via a Receiver Operating Characteristic (ROC) formulation. The algorithm was employed on a large dataset of over 100 magnetic observatories in California totaling hundreds of thousands of station-days. Using the ROC curve to evaluate its performance, this implementation of the algorithm obtained a 2.20 z-score. This number improved with the preliminary attempt at removing a severe cultural noise source. This work emphasizes an analytic framework more than parametric exploration or optimization, nevertheless there appears to be some suggestion of predictive power in the magnetic field time series.