Abstract
In recent decades, ecological forecasting, i.e. the prediction of natural systems in time, has become a critical aspect of ecology. Indeed, ongoing global challenges such as climate change, environmental pollution and biodiversity loss have rendered the proficient forecasting of ecosystem states, functioning and services a necessity. Yet, most ecological studies focus solely on explanatory analyses, deepening our understanding of natural systems but failing to validate their insights through the use of forecasts and to inform possible stakeholders by predicting their study system into the future. Moreover, due to the considerable complexity of nature, ecology is frequently perceived as not being a predictive science, a view that might discourage ecologists from attempting to forecast nature. As a consequence of the low prevalence of forecasts in ecological research, little is known about which biological properties influence the quality of forecasts and whether there are some general indicators of ecological forecast skill. In short, currently there is a lack of studies investigating the drivers of forecast skill.
In this thesis, to advance the field of ecological forecasting I investigated how forecasts of species abundances and community biomass were influenced by several system properties and environmental variables. Across four research chapters, the potential drivers that I studied were: species traits (body size and growth rates), inter-species properties (number and strength of species interactions, producer defensive strategies against stress), time series properties (temporal stability, autocorrelation, permutation entropy), community properties (species richness and evenness, number of functional groups), environmental factors (fluctuating temperature, decreasing light, multiple stressors: temperature, salinity and resources), and sampling design and modeling properties (number and frequency of samplings, number of used predictors). To do so, I used long-term experiments with microbial aquatic communities and real-world lake plankton dynamics, and I employed various non-mechanistic forecasting approaches and one semi-mechanistic modeling framework.
I found that most of the tested variables influenced forecasts. For instance, I forecasted species with more but on average weak interactions better than species with few but strong interactions. Forecasts were also better for species with smaller growth rates and, in constant environments, they were also better for less speciose systems. However, this last result depended on the environmental setting, as decreasing light conditions inverted the relationship between biodiversity and forecast skill so that I forecasted communities with more species better. Further, environmental change, in the form of multiple stressors, also resulted in a reduction of forecast skill. Together, these results indicate that complexity not necessarily hinders forecasting, as in some cases the opposite is true (possibly because the present drivers led to a simplification of the community dynamics). Moreover, this increased knowledge of what determines how well we can forecast has the potential to translate to better forecasts in general, which in turn should lead to better-informed decisions in ecosystem conservation and management.
In contrast, I also encountered a case in which including promising drivers of forecast skill (i.e. species defenses against stress) into a semi-mechanistic model did not result in better biomass forecasts, even though the explanatory performance starkly improved. This result exemplifies the often-overlooked fact that good explanations do not necessarily imply good predictions. Therefore, this result also highlights the need for more frequent out-of-sample validation in ecology.
Lastly, I also show that forecasts generally benefited from an increase in sampling effort, suggesting that many systems might not be monitored at the optimal frequency. I further used this insight to exemplify how forecast quality can be used a guide to find the best sampling design and how this can improve tasks not limited to forecasting (for instance by bettering the estimation of species interactions). Therefore, this result highlights how forecasts can be used to advance ecological inference and understanding.
In conclusion, forecasts serve the purposes of advancing and validating ecological theory as well as informing decision makers and stakeholders about the future states and services of the systems of interest. The improvements in our knowledge about ecological forecasting and its drivers presented in this thesis could lead to overall better forecasts and therefore could also help with the mentioned purposes of ecological forecasting. Moving forward, I think that forecasting should become more prominent in ecology. This can be achieved through a greater focus on it in ecological education and through more incentives to forecast (provided, for instance, by funding agencies, scientific journals and forecasting challenges). In the end, it is my hope that this thesis successfully shows the importance of ecological forecasting and that it helps dispelling the myth of the unpredictability of ecology due to its complexity.