Improved Automatic Seismic Bulletins Via Likelihood-Based Model Fit Scores for Classification

Automatically produced seismic bulletins, such as the Standard Event List (SEL1) of the International Data Centre (IDC), contain many malformed events. Analysts frequently reject, split, or substantially modify events before they enter refined lists such as the Late Event Bulletin (LEB). A key weakness is that standard pipelines rely almost exclusively on positive detections and ignore non-detecting stations, even though a station that should have detected an event but did not provides strong evidence against that event’s legitimacy. We address this by creating an event scoring function that incorporates informative missing data from non-detections and other prior seismic knowledge. Using LEB as a training reference, we model for each station the detection probability and distribution of key observed summaries. For every SEL1 event, we compute likelihood-based model-fit score features that quantify how well the candidate event explains the observed detection pattern across the network. These physics-driven features feed a simple classifier that outputs a legitimacy score for each event. Applied to one year of independent SEL1 test data, the classifier identifies more than 72\% of false events while falsely flagging only 5\% of valid ones. Low-scoring events retained by analysts often correspond to cases where the score helps diagnose and correct data problems. The transparency and interpretability of the classifier make it well-suited to seismic monitoring, potentially as an analyst support tool, while preserving the benefits of a high-performing machine-learning model.