Model-free forecasting outperforms the correct mechanistic model for simulated and experimental data

Abstract

Accurate predictions of species abundance remain one of the most vexing challenges in ecology. This observation is perhaps unsurprising, because population dynamics are often strongly forced and highly nonlinear. Recently, however, numerous statistical techniques have been proposed for fitting highly parameterized mechanistic models to complex time series, potentially providing the machinery necessary for generating useful predictions. Alternatively, there is a wide variety of comparatively simple model-free forecasting methods that could be used to predict abundance. Here we pose a rather conservative challenge and ask whether a correctly specified mechanistic model, fit with commonly used statistical techniques, can provide better forecasts than simple model-free methods for ecological systems with noisy nonlinear dynamics. Using four different control models and seven experimental time series of flour beetles, we found that Markov chain Monte Carlo procedures for fitting mechanistic models often converged on best-fit parameterizations far different from the known parameters. As a result, the correctly specified models provided inaccurate forecasts and incorrect inferences. In contrast, a model-free method based on state-space reconstruction gave the most accurate short-term forecasts, even while using only a single time series from the multivariate system. Considering the recent push for ecosystem-based management and the increasing call for ecological predictions, our results suggest that a flexible model-free approach may be the most promising way forward. U nderstanding fluctuations in species abundance is a longstanding goal of ecology and is particularly important for the conservation of severely depleted populations (1, 2). Dramatic changes in species abundance are common (3) and large declines can have disastrous impacts on ecosystem users (4). Accurate forecasts could allow for improved conservation efforts and increased fishery yields, however, ecological surprises are common (5), and accurate predictions remain a major challenge (6).As ecological time series continue to lengthen, and computerintensive statistics become more convenient, there has been an increasing trend toward fitting complex mechanistic models that incorporate both process and observation error (i.e., state-space models) (7). A Bayesian fitting procedure for such a model typically involves computationally intensive Markov chain Monte Carlo (MCMC) sampling of the joint posterior probability distribution to find the best fitting parameters (8). Previous work concluded that misspecified mechanistic models can produce poor forecasts (9), but it has been suggested that state-space models may be more robust to misspecification due to their inclusion of a process-error term (8); however, their forecast accuracy remains largely untested.Model-free time series analysis provides an alternative method for generating forecasts that, unlike most state-space fitting procedures, does not require intensive computations. Although there exists a vas...

Elevated nonlinearity as an indicator of shifts in the dynamics of populations under stress

“…In either case, how well the attractor is reconstructed can be verified by its ability to forecast future states. In fact, EDM has been shown to outperform equation-based approaches at forecasting recruitment in sockeye salmon populations [1], dynamics of Pacific sardines [29] and the fate of experimental flour beetle populations [30]. With its capacity to forecast the future state of a system, EDM could potentially be a useful alternative approach for detecting the proximity of a system to nearby bifurcations.…”

Equation-free mechanistic ecosystem forecasting using empirical dynamic modeling

“…In the standard parametric approach, which implicitly assumes that the selected model and its equations are essentially correct, the equations (really just mechanistic hypotheses) can lack the flexibility to describe the nonlinear dynamics that occur in nature. Consequently, these parametric models tend to perform poorly as descriptions of reality, with little explanatory or predictive power (2,3), and limited usefulness for prediction and management.…”

A fast universal self-tuned sampler within Gibbs sampling

“…The FUSS algorithms are also able to overcome severe computational issues in some statistical frameworks where other, even more sophisticated, MCMC techniques seem to fail [39,40]. In order to show this capability, we address the estimation of fixed parameters of a chaotic system, which is considered a very challenging problem in the literature [39][40][41].…”

“…In order to show this capability, we address the estimation of fixed parameters of a chaotic system, which is considered a very challenging problem in the literature [39][40][41]. This type of systems are often utilized for modeling the evolution of population sizes, for instance in ecology [39]. Let us consider a logistic map [42] perturbed by multiplicative noise,…”