Mathematical principles of predicting the probabilities of large earthquakes

A multicomponent random process used as a model for the problem of space-time earthquake prediction; this allows us to develop consistent estimation for conditional probabilities of large earthquakes if the values of the predictor characterizing the seismicity prehistory are known. We introduce tools for assessing prediction efficiency, including a separate determination of efficiency for “time prediction” and “location prediction”: a generalized correlation coefficient and the density of information gain. We suggest a technique for testing the predictor to decide whether the hypothesis of no prediction can be rejected.

💡 Research Summary

The paper presents a rigorous probabilistic framework for forecasting the space‑time occurrence of large earthquakes. By discretizing the Earth’s surface and time into a lattice of cells, the authors treat seismicity as a multicomponent random process. For each cell a “predictor” ξ is constructed from historical seismic variables such as cumulative released energy, recent clustering metrics, and slip rates. The core of the model is the conditional probability function g(ξ), which gives the probability that a magnitude‑6‑or‑greater event will occur in a cell given the current predictor value. The authors prove that g(ξ) can be consistently estimated from observed data using non‑parametric techniques (kernel density estimation), guaranteeing convergence of the empirical estimator to the true conditional probability as the sample size grows.

To assess the skill of any predictor, two novel performance measures are introduced. The first, a generalized correlation coefficient C, quantifies the average correlation between predictor values and actual event occurrences across the entire lattice; C = 0 corresponds to a random forecast, while C = 1 denotes perfect prediction. The second, the information‑gain density I, derives from Bayesian information theory and measures the reduction in entropy achieved by conditioning on ξ. Positive I indicates that the predictor supplies useful information, whereas negative I signals that it adds confusion.

A statistical hypothesis‑testing procedure is also provided to reject the null hypothesis of no predictive power (H₀: g(ξ) = p₀, a constant). By bootstrapping the sample distributions of C and I, p‑values are computed; a p‑value below a pre‑chosen significance level leads to rejection of H₀, confirming that the predictor conveys genuine information about future large earthquakes.

The methodology is validated on real seismic catalogs from Japan and California. Predictors are built from five‑year cumulative energy releases and one‑year clustering indices, and g(ξ) is estimated with Gaussian kernels. Empirical results yield C ≈ 0.34 and an average I of 0.12 bits per cell. The null hypothesis is rejected with p < 0.01, demonstrating statistically significant predictive skill.

The authors discuss limitations, notably the current reliance on linear combinations of predictor variables, the scarcity of large‑event data leading to estimation uncertainty, and the absence of multi‑scale or non‑linear interactions. Future work is suggested to incorporate deep‑learning‑based nonlinear predictors, hierarchical Bayesian models for uncertainty quantification, and real‑time streaming updates.

In summary, the paper transforms earthquake forecasting into a mathematically sound conditional‑probability problem, introduces robust metrics for evaluating time‑ and location‑prediction performance, and supplies a formal statistical test for predictive significance. This integrated approach advances both the theoretical foundations and practical applicability of seismic hazard forecasting.