COS 72-4 - CANCELLED - Environmental variability, not intrinsic predictability, controls the forecastability of daily cyanobacterial dynamics over ten summers

Abstract: Forecasting messy, real-world ecological data is a major challenge. The forecastability of a time series is a function of both its intrinsic predictability, i.e., a measure of the inherent stochasticity of the data, and its realized predictability, i.e., the ability of a model to predict future dynamics. Previous work has shown that intrinsic predictability is an important driver of realized predictability among ecological variables, but it remains unknown if the importance of intrinsic predictability changes over time within an ecological time series. To identify the limits of ecological predictability, we need to know: 1) How dynamic is the intrinsic predictability of an ecological variable over time? and 2) Does the role of intrinsic predictability as a predictor of realized predictability vary in different environmental conditions? We answered these questions by analyzing a unique dataset of daily cyanobacterial densities measured in a temperate lake over ten sequential summers. First, we quantified intrinsic predictability separately for each summer as permutation entropy, an information theory-based metric of the complexity of the daily cyanobacterial dataset. Second, we constructed empirical dynamic models that were trained on the first 80% of each summer’s time series and then forecasted the remaining 20% of the summer time series. We quantified realized predictability for each summer using multiple goodness-of-fit metrics from the out-of-sample evaluation of the forecasted data vs. observations.



We observed substantial variability in intrinsic and realized predictability among summers. In general, permutation entropy of the daily cyanobacterial time series data was high (indicating low intrinsic predictability), with values ranging by more than 33% among the ten summers. In contrast, realized predictability was extremely variable among years (e.g., R² ranged from 0-0.61). Surprisingly, we did not observe a significant association between intrinsic predictability and realized predictability across the summers; rather, environmental variability was a more important predictor of realized predictability than intrinsic predictability. Over the ten years, summers with stronger thermal stratification exhibited lower realized predictability. These strongly thermally-stratified summers were associated with lower cyanobacterial densities, which may be due to lower recruitment rates of cyanobacteria from the sediments to the water column. Our data suggest that the forecastability of cyanobacteria may decrease as the strength of lake thermal stratification increases with warming air temperatures. Moreover, our study provides valuable insight on the limits of cyanobacterial predictability, as quantifying both intrinsic and realized predictability informs how much further predictability can increase with model and data improvements.