Long-Term Probabilistic Forecast for M ≥ 5.0 Earthquakes in Iran

In this study, a long-term forecasting model is proposed to evaluate the probabilities of forthcoming M >= 5.0 earthquakes on a 0.2 degrees grid for an area including the Iranian plateau. The model is built basically from smoothing the locations of preceding events, assuming a spatially heterogeneous and temporally homogeneous Poisson point process for seismicity. In order to calculate the expectations, the space distribution, from adaptively smoothed seismicity, has been scaled in time and magnitude by average number of events over a 5-year forecasting horizon and a tapered magnitude distribution, respectively. The model has been adjusted and applied considering two earthquake datasets: a regional unified catalog (MB14) and a global catalog (ISC). Only the events with M >= 4.5 have been retained from the datasets, based on preliminary completeness data analysis. A set of experiments has been carried out, testing different options in the model application, and the average probability gains for target earthquakes have been estimated. By optimizing the model parameters, which leads to increase of the predictive power of the model, it is shown that a declustered catalog has an advantage over a nondeclustered one, and a low-magnitude threshold of a learning catalog can be preferred to a larger one. In order to examine the significance of the model results at 95% confidence level, a set of retrospective tests, namely, the L test, the N test, the R test, and the error diagram test, has been performed considering 13 target time windows. The error diagram test shows that the forecast results, obtained for both the two input catalogs, mostly fall outside the 5% critical region that is related to results from a random guess. The L test and the N test could not reject the model for most of the time intervals (i.e. similar to 85 and similar to 62% of times for the ISC and MB14 forecasts, respectively). Furthermore, after backwards extending the time span of the learning catalogs and repeating the L test and N test for the new dataset as well as the R test, it is observed that a low-quality longer catalog does not essentially improve the predictive skill of the model than a high-quality shorter one. The performed retrospective tests suggest that the model yields some statistically acceptable efficiency for the studied area, at least with respect to the spatially uniform reference model. Thus, the considered model may provide useful information as a reference for more refined time-independent models and also in combination with long-term indications from seismic hazard maps; this is particularly relevant in areas characterized by a high level of predicted ground shaking and high forecast rate.