Physics-based and statistical earthquake forecasting in a continental rift zone: the case study of Corinth Gulf (Greece)

I perform a retrospective forecast experiment in the most rapid extensive continental rift worldwide, the western Corinth Gulf (wCG, Greece), aiming to predict shallow seismicity (depth < 15 km) with magnitude M >= 3.0 for the time period between 1995 and 2013. I compare two short-term earthquake clustering models, based on epidemic-type aftershock sequence (ETAS) statistics, four physics-based (CRS) models, combining static stress change estimations and the rate-and-state laboratory law and one hybrid model. For the latter models, I incorporate the stress changes imparted from 31 earthquakes with magnitude M >= 4.5 at the extended area of wCG. Special attention is given on the 3-D representation of active faults, acting as potential receiver planes for the estimation of static stress changes. I use reference seismicity between 1990 and 1995, corresponding to the learning phase of physics-based models, and I evaluate the forecasts for six months following the 1995 M = 6.4 Aigio earthquake using log-likelihood performance metrics. For the ETAS realizations, I use seismic events with magnitude M >= 2.5 within daily update intervals to enhance their predictive power. For assessing the role of background seismicity, I implement a stochastic reconstruction (aka declustering) aiming to answer whether M > 4.5 earthquakes correspond to spontaneous events and identify, if possible, different triggering characteristics between aftershock sequences and swarm-type seismicity periods. I find that: (1) ETAS models outperform CRS models in most time intervals achieving very low rejection ratio R-N = 6 per cent, when I test their efficiency to forecast the total number of events inside the study area, (2) the best rejection ratio for CRS models reaches R-N = 17 per cent, when I use varying target depths and receiver plane geometry, (3) 75 per cent of the 1995 Aigio aftershocks that occurred within the first month can be explained by static stress changes, (4) highly variable performance on behalf of both statistical and physical models is suggested by large confidence intervals of information gain per earthquake and (5) generic ETAS models can adequately predict the temporal evolution of seismicity during swarms. Furthermore, stochastic reconstruction of seismicity makes possible the identification of different triggering processes between specific seismic crises (2001, 200304, 2006-07) and the 1995 aftershock sequence. I find that: (1) seismic events with M >= 5.0 are not a part of a preceding earthquake cascade, since they are characterized by high probability being a background event (average P-back > 0.8) and (2) triggered seismicity within swarms is characterized by lower event productivity when compared with the corresponding value during aftershock sequences. I conclude that physics-based models contribute on the determination of the 'new-normal' seismicity rate at longer time intervals and that their joint implementation with statistical models is beneficial for future operational forecast systems.