Investigating the predictive power of seismic statistical features using ensemble learning

Earthquake prediction is an extremely challenging problem, one that has been in the past (and sometimes still is) claimed to be impossible. Given this undisputed high level of difficulty, work that reports a high level of prediction success might reasonably be regarded with a degree of caution. We will discuss here how these results may in many cases be due to data leakage. However, a recent paper co-authored by one of us has shown a promising level of predictive ability even when its methodology strictly controls for possible overfitting and data leakage. We here build on that prior work by asking if the demonstrated predictive value of the seismic statistical features used there is due to their being able to capture domain-specific knowledge. Specifically, we compare the value of the same set of 60 seismic statistical features used in the aforementioned previous work to the value of a set of 428 generic time series features from the tsfresh package. We train an XGBoost model to predict if there will be an earthquake of magnitude M ≥ 5 in the following 15 days, and find models using the seismic statistical features can attain AUCs of up to 0.87, while models using the tsfresh features alone cannot obtain results substantially better than random. It therefore does appear that seismic-specific catalogue features are able to capture valuable information about subsurface conditions prior to an impending earthquake. We do not attempt to carry out operational earthquake prediction, considering it premature at this time. However, the demonstrated seismic-specific origin of the predictive power of our features gives hope that by augmenting and enhancing them such prediction may become feasible, and we conclude by discussing some novel directions for future work.