The development of earthquake probability forecasts from a hidden Markov model: An example from the Killini Region, Greece

Technical Abstract

A four-state hidden Markov model is used to develop 1-day probability forecasts for earthquakes in the Killini region of Greece. The model allows the rates in each state, and the transition probabilities between states, to be estimated from the earthquake occurrence data. Forecasts proceed by using the model to estimate the current state probabilities, and using these to estimate the forecast probabilities for the next 1-day period. The data used is provided by the National Observatory in Athens, and comprises all events over local magnitude 3.2 in the Killini seismic zone. The forecasts are prepared initially for the region as a whole, and then partitioned out over a spatial grid covering the observation region, and into magnitude classes. The final results therefore consist of daily forecasts for each grid in the spatial region and for each magnitude class. The forecasts from the hidden Markov model are compared to forecasts obtained from applying a temporal ETAS (Epidemic Type Aftershock Sequence) model to the same data and region. The hidden Markov model appears to have advantages when the data is clustered, and the clusters only partially follow the traditional main-shock, aftershock pattern.