Integrated Seismic Risk Modelling and Parametric Disaster Financing in Indonesia Using Probabilistic Hazard Analysis, Machine Learning, and Actuarial Simulation.

This study examines whether combining machine-learning models with seismicity indicators can improve the calibration and skill of short-term earthquake probabilities for parametric disaster financing in Indonesia. An integrated framework is developed that combines probabilistic seismic hazard analysis (PSHA), machine learning, and actuarial simulation. Using the United States Geological Survey (USGS) catalog for 1925–2025, Gutenberg–Richter (GR) parameters and long-term Poisson exceedance rates are estimated for events with M ≥ 5.0 on a 0.5° × 0.5° grid. These features, together with geographic coordinates, are used as inputs to tree-based models that predict the probability of at least one event with M ≥ 5.0 occurring within the next five years. In walk-forward validation, Random Forest and Gradient Boosting achieve modest discrimination (Area Under the Receiver Operating Characteristic Curve (AUROC) ≈ 0.60) but yield small yet well-calibrated improvements over the GR–Poisson baseline, with lower Brier scores and more reliable probabilities. For the financial component, a parametric payout rule is applied in which payouts increase linearly with magnitude between 5.0 and 7.5. Monte Carlo simulations based on the long-term rates yield estimates of expected annual loss and the 95% tail value-at-risk (TVaR), highlighting skewed risk concentrated along the Sumatra subduction corridor. The framework assumes independent Poisson occurrence within each grid cell, uses a coarse spatial resolution, and does not explicitly model spatial correlation or aftershock clustering. [ABSTRACT FROM AUTHOR]