Integrating Transfer Learning and Socio-Economic Value Metrics to Improve Eruption Forecasting and Decision-Making at Data-Limited Volcanoes

Forecasting volcanic eruptions remains challenging due to the scarcity of long-term monitoring data, the diversity of volcanic systems, and the difficulty of distinguishing subtle precursory signals from background variability. Here we proposed two methodological advances that offer complementary pathways to improve both scientific skill and operational decision-making: (i) machine-learning transfer forecasting based on ergodic seismic precursors, and (ii) socio-economic valuation of forecasts using the Potential Economic Value (PEV) framework.

First, we show that seismic precursors exhibit ergodic behavior, enabling machine-learning models trained on multi-volcano datasets to forecast eruptions at completely unseen, data-limited volcanoes. Using 73 years of continuous seismic data from 24 volcanoes and 41 eruptions, transfer-learning models identify statistically recurrent time-series features that strengthen prior to eruptions and can be effectively transferred between systems with distinct eruptive characteristics. Out-of-sample tests show forecasting skill comparable to tailored local models and exceeding benchmarks based on seismic amplitude. These results indicate that cross-volcano precursor patterns can provide robust forecasting capability even where local eruption histories are sparse, supporting global applicability of generalized forecasting tools.

However, forecast skill alone does not guarantee societal value. To address this gap, we introduce the potential economic values (PEV) framework to quantify the operational benefits of these forecasts by balancing the manageable costs of false alarms against the catastrophic consequences of missed eruptions. Retrospective analyses at Whakaari (2019) and Ontake (2014), combined with hypothetical high-impact scenarios, shows that even imperfect ML forecasts can reduce avoidable losses by 30–90%. PEV reveals that forecast value is maximized not by optimizing statistical accuracy, but by minimizing missed eruptions—highlighting the asymmetric socio-economic impacts of forecast errors. Optimal operational thresholds emerge within a stable range across volcanoes and cost assumptions, underscoring transferability of the framework.

By combining cross-volcano transfer learning with cost-based evaluation, our integrated framework advances two frontiers in volcanic hazard science: (1) improving eruption forecasting capability at data-limited volcanoes using ergodic precursor patterns, and (2) enabling monitoring agencies to select operational thresholds that maximize societal benefit rather than statistical performance alone. This approach supports more transparent, defensible, and economically efficient decision-making during volcanic unrest and provides a scalable pathway toward next-generation, globally transferable hazard-forecasting systems.