Machine-Learning Forecasting of Strong Aftershocks in New Zealand Using NESTORE: Insights from Two Testing Approaches

Located along the active boundary between the Australian and Pacific plates, New Zealand is subject to widespread deformation and high seismicity, with major earthquakes reaching magnitudes up to 7.8. This tectonic setting makes the region well suited for evaluating operational, data-driven forecasting tools such as NESTORE. Improving the understanding and forecasting of seismic activity is essential for rapid hazard assessment, risk evaluation, and mitigation planning. We therefore applied the machine-learning–based probabilistic forecasting algorithm NESTORE (NExt STRong Related Earthquake, Gentili et al., 2023) to New Zealand’s seismicity. NESTORE evaluates nine features related to aftershock occurrence, source area evolution, and temporal trends in magnitude and radiated energy, calculated over progressively longer time windows after each mainshock. Using these features, the algorithm estimates the probability that a mainshock of magnitude Mm will be followed by another event of magnitude ≥ Mm–1 within the spatial and temporal bounds of its seismic cluster. Clusters that experience such a strong aftershock are labelled “Type A,” indicating higher potential hazard, while others are labelled “Type B.” For each cluster, the algorithm returns the probability of belonging to Type A. NESTORE’s performance was assessed using two approaches. The first was chronological: a cutoff time was selected so that the model was trained on clusters occurring before this cutoff, and then applied to forecast the behaviour of clusters occurring afterwards. The second approach used stratified k-fold cross-validation (Zeng et al., 2000) to test generalisation across multiple randomised partitions of the dataset. To further strengthen model training, we incorporated the outlier-detection procedure REPENESE (RElevant features, PErcentage class weighting, NEighbourhood detection and SElection, Gentili et al. 2025). Our findings indicate that k-fold cross-validation provides more stable and reliable performance than the chronological approach, although changes in the catalogue may make the newer clusters a more valuable test set. NESTORE correctly classified 88% of seismic clusters 18 hours after the mainshock, including 77% of Type A and 92% of Type B clusters. Importantly, the algorithm successfully identified the Canterbury/Christchurch 2010–2011 sequence – a major, destructive Type A cluster.