Ecosystem forecasts often aim toward forecasting the distribution, occurrence, or abundance of particular species. Yet there are many cases where it is the interaction between members of a species and humans that is the question of concern. Examples include harmful or toxic species, car-wildlife collisions, harvesting, and interactions between threatened species and human activities. There can be multiple processes and layers of uncertainty between the distribution of a species and its impact on humans. This raises the question: How can we forecast human-wildlife interactions directly? We built a forecasting program that uses reports of human-wildlife interaction—mainly citizen reports of species—as a basis for building near-term ecological forecasts. The system uses species distribution models in an adaptive/learning mode and evaluates forecasts based on predictions of future human-wildlife interactions. We tested the system on a range of events of interest, including car-moose collisions, jellyfish outbreaks, recreational fishing, and tick encounters, among others.
Results/Conclusions
The system performed well in a variety of forecasting contexts based on forecast skill metrics. One advantage to this approach is that forecasting systems can be spun up quickly and in some cases reach reliable levels within a few weeks of collecting data, even in cases where a historical data baseline does not exists. However, because of the empirical nature of the forecasting method, forecast performance and range depends on the nature of autocorrelation in the system. There also can be bias feedback loops, depending on the structure of communication within the reporting-forecasting system. Despite potential pitfalls, the approach provides an alternative that overcomes some of the drawbacks of conventional forecasting approaches.