Evaluation of earthquake stochastic models based on their real-Time forecasts: A case study of Kaikoura 2016

Abstract

The M7.8 Kaikoura New Zealand earthquake started at 2016-11-13 11:02:56 (UTC) with epicentre (173.02°E, 42.69°S), 15 km NE of Culverden, and lasted for about 2 min. It caused multiple fault ruptures to the north as far as Seddon, the location of a large sequence in 2013. Since the mainshock, the bulk of the aftershock activity has also migrated to the north. We analyse real-Time probability forecasts produced during the Kaikoura 2016 aftershock sequence based on a spatial Epidemic Type Aftershock Sequence (ETAS) model. Forecasts were derived by simulating the model forward over the required time interval multiple times. Each forecast was evaluated at the end of the forecast time interval by comparing with the number of events that eventually occurred. Further, the spatial and temporal forecast characteristics were evaluated by comparing the actual log-likelihood with those of the simulations. We show that the model was forecasting too few aftershocks immediately after the mainshock, and too many aftershocks in the later stages of the sequence. The too few aftershocks are probably caused by many missing smaller events early in the sequence and an initial large underestimate of the mainshock magnitude, being 6.6 with a final solution of 7.8 three days later. Overforecasting in the later stages is partially caused by a magnitude discrepancy between the CUSP and SeisComP3 software. The CUSP software was used to solve event solutions until the end of 2011, and the SeisComP3 software has been used since 2012. Our ETAS model parameters were estimated using catalogue data until the end of 2011 (CUSP). Other catalogue, model and methodological problems become evident during such a real-Time exercise and these are also discussed.