The forecasted mean of 80 percent of wild populations and communities shows no change, and why

Broad-scale shifts in biodiversity demand a deeper understanding of both historical and future time trends across the Tree of Life. Retrospective analyses indicate that some groups, such as birds, fishes and insects, have experienced significant declines in recent decades. Broader investigations, on the other hand, reveal a more complex picture, with declines and increases roughly balanced. Here, we address the prospective task to statistically forecast biodiversity trends and assess our ability to do so based on available data. We analyzed approximately 43,000 population-level and 10,000 community-level time series from localities across the globe and covering taxa across the Tree of Life. Using unobserved components models and model selection, we found that only 23.5% of wild populations exhibit forecasted time trends (increasing and decreasing ones roughly balanced), while 76.5% show a forecasted constant mean. For communities (temporal - and {beta}-diversity), this figure rises to 84.0%. To explore the underlying reasons for the high prevalence of forecasted constant means, we used machine learning techniques (random forest models and SHAP analysis). We identified high variation in the rate of change (RoC) in the time series as the most important predictor of a forecasted constant mean. Further, life-history traits and locality accounted for 39.6% of the variation in the RoC variance of anmiotes, and sampling unit explained 44.4% of that variability for dicot plants. In summary, despite time trends being identifiable in ecological time-series data, parsimonious statistical models will very often suggest a forecasted constant mean. Our results offer a tempered perspective that can help realistically inform stakeholders and the general public, and guide the assessment of conservation action outcomes.