Fig. 4.

Forecasting community-level dynamics based on attractor reconstruction. a Community-level forecasting. Predicted and observed community structure is linked for each day on the axes of NMDS (prediction based on S-map with optimized θ; seven-day-ahead forecasting). Results of Soil/Medium-A treatment are shown: see Additional files 10 and 11: Figures S10–11 for full results). b Factors explaining variation in community-level prediction results. A generalized linear mixed model (GLMM) of dissimilarity between predicted and observed community structure was constructed (one-day-ahead forecasting). c Detailed comparison of nonlinear and linear forecasting approaches. S-map results with optimized nonlinearity parameter were compared with results of S-map assuming linear dynamics for all ASVs. See Additional file 12: Figure S12 for full results