Comment on “Potential short-term earthquake forecasting by farm animal monitoring” by Wikelski, Mueller, Scocco, Catorci, Desinov, Belyaev, Keim, Pohlmeier, Fechteler, and Mai
Summary
WIVVI summary (paraphrased): The authors independently reanalyzed the farm-animal activity and seismic data used by Wikelski and colleagues to propose potential short-term earthquake forecasting. They argued that the original analysis considered occasions when seismic anomalies were preceded by animal anomalies but did not adequately count animal anomalies followed by no seismic event or seismic anomalies without preceding animal activity. Using Molchan error diagrams and randomized anomaly sequences, they found that the observed animal-activity patterns could not be distinguished from random patterns and did not demonstrate forecasting skill. The study concludes that the available dataset does not justify using the observed farm-animal anomalies for earthquake forecasting.
Source
Ethology
Keywords
farm animal behavior, earthquake forecasting, animal activity anomalies, Molchan diagram, forecasting skill, false alarms, failures to predict, randomization analysis, seismic clustering, Norcia earthquake
Source Type
commentary
Key Findings
When successful predictions, false alarms, and failures to predict were all included, the animal-based forecasting results lay close to random-performance expectations. Many randomized animal-anomaly sequences performed better than the observed sequence. The proposed 20-hour alarm window left the alarm active for much of the monitoring period, making its apparent success operationally uninformative. A simple forecast based only on recent seismic activity generally performed better because earthquakes cluster in time. Randomization tests also showed that the reported relationship between anticipation time and earthquake distance could arise from the space-time clustering of the earthquakes. The reanalysis therefore found no demonstrated forecasting power in the animal-activity signal.
Limitations
This was a statistical reanalysis of one previously published dataset rather than an independent animal-monitoring experiment. The underlying observations involved one group of cows, sheep, and dogs in one earthquake-prone region during three relatively short monitoring periods. Animal activity contained substantial non-seismic variation associated with daily routines, farming activities, housing conditions, weather, and other disturbances. The earthquake sequence was strongly clustered, reducing the number of effectively independent seismic observations. Consequently, the analysis directly challenges the forecasting interpretation of this dataset but cannot establish that animals never respond to any physical process associated with an impending earthquake. Longer prospective studies across independent locations, species, and earthquake sequences would be required.
Methods Summary
The authors downloaded the animal-activity data, earthquake-derived peak-ground-acceleration series, and code supplied with the original study. The dataset covered three monitoring periods involving tagged cows, sheep, and dogs near the 2016 magnitude 6.6 Norcia earthquake sequence in Italy. Animal and seismic anomalies were defined using the original two-standard-deviation thresholds. Molchan error diagrams were constructed for alarm windows of different durations so that successful forecasts, false alarms, and failures to predict were evaluated together. Forecast performance was compared with randomized timings of animal anomalies and with a simple forecast based on the temporal clustering of seismic activity. The reported relationship between anticipation time and hypocentral distance was also tested using 1,000 randomized animal-anomaly sequences.