Investigating the predictive power of seismic statistical features using ensemble learning.

Area of Science:

• Geophysics
• Seismology
• Data Science

Background:

• Earthquake prediction is a complex challenge, often met with skepticism.
• Previous studies sometimes suffer from data leakage, inflating success rates.
• A prior study demonstrated predictive ability using seismic features while controlling for data leakage.

Purpose of the Study:

• To determine if seismic statistical features capture domain-specific knowledge for earthquake prediction.
• To compare the predictive power of seismic features against generic time series features.
• To validate the source of predictive information in earthquake forecasting.

Main Methods:

• Compared 60 seismic statistical features against 428 generic time series features from the tsfresh package.
• Utilized an XGBoost model for predicting earthquakes (magnitude M ≥ 5) within a 15-day window.
• Employed rigorous methodology to prevent overfitting and data leakage.

Main Results:

• Models using seismic statistical features achieved an Area Under the Curve (AUC) up to 0.87.
• Models using only tsfresh generic features performed no better than random chance.
• This highlights the significance of seismic-specific data in capturing pre-earthquake subsurface information.

Conclusions:

• Seismic-specific features demonstrably capture valuable information for earthquake prediction.
• Generic time series features lack the domain-specific insight required for effective forecasting.
• Future research can enhance these seismic features for potential operational earthquake prediction.