Abstract
Forecasting the likely style and chronology of activity within an eruption is a complex issue that has received far less attention than forecasting the onset and/or the magnitude. By developing a global data set of coded phases (discrete styles of activity within previous eruptions), we can model the resulting data using a semi-Markov chain. Given enough data, we were able to examine the question of whether analogue-based strategies for subsetting the data can improve forecasting performance of phase chronology and style within ongoing eruptions. This work required inclusion of a “null analogue” element to ensure no surprises, that is, phase transitions or durations that were not in the data set and hence cannot be predicted. We have significantly expanded, and made available, our curated data set on eruption phases, which now contains 2670 eruptions (6871 phases), of which 56% are multi-phase. This increases the data set by 283% and includes 95% of Holocene eruptions with text descriptions. We find that, with the notable exception of shields, limiting the analogue set on the basis of volcano morphology and/or composition is not significantly more informative than using the entire data set. Dynamically adjusting the data limits by eliminating eruptions without the observed phase as the eruption progresses provided little benefit, although subsetting on the basis of VEI may have some utility. At the individual volcano level, non-analogue models can outperform the entire data set, if the target volcano has relatively unique behavior and/or a large enough record of phased eruptions.
Keywords
Eruption forecasting, eruption phases data set, intra-eruption chronology, multi-phase eruptions, volcano and eruption analogs