Ecological forecasting

Ecological forecasting: how predictable is ecology?

Ecology has traditionally been very good at explaining why ecological systems behaved as they did. A more demanding question is whether ecological understanding allows us to say what they will do next. That distinction lies at the heart of ecological forecasting. Rather than asking only whether a model fits past observations, forecasting evaluates predictions against observations that were not used to construct the prediction.

Work by Owen Petchey, Frank Pennekamp and collaborators has approached forecasting from a particularly fundamental direction. Instead of asking only which statistical or mechanistic model produces the best forecast, they have asked: how predictable are ecological systems in the first place, and what determines that predictability?

The ecological forecast horizon

A starting point was the concept of the ecological forecast horizon, developed by Petchey and a large group of collaborators in 2015 [1]. The basic idea is intuitive. A weather forecast may be useful several days ahead but progressively less informative further into the future. Ecological forecasts should similarly have a distance beyond which their usefulness falls below an acceptable threshold.

Petchey and colleagues defined the forecast horizon as the distance over which a useful ecological forecast can be made [1]. Most obviously, that distance can be time: how many days, months or years ahead can population abundance be predicted? But the concept is deliberately broader. Forecast horizons can also exist across space, environmental conditions such as temperature, or even biological differences among organisms. The framework therefore shifts attention from asking whether ecology is “predictable” in general to a more quantitative question: How far can we predict, with what accuracy, and what determines that limit?

Why forecasts fail

There are many reasons why ecological forecasts deteriorate. Environmental conditions become increasingly uncertain further into the future. Population dynamics contain demographic stochasticity. Parameters are estimated imperfectly. Species interact. Evolution can change the properties of organisms themselves.

The forecast-horizon framework makes these sources of uncertainty experimentally and quantitatively interesting because each may shorten the distance over which predictions remain useful [1]. This perspective also suggests that predictability should itself be treated as a property to study rather than simply a measure of whether a particular model worked.

Is the system unpredictable, or is the model poor?

This distinction motivated subsequent work on intrinsic predictability. Suppose a forecasting model performs badly. There are at least two very different explanations. The ecological system may contain so much stochastic or chaotic variation that accurate forecasting is intrinsically difficult. Alternatively, the system may be quite predictable, but the model—or the available data—may be inadequate.

Pennekamp and colleagues explored whether information-theoretic properties of ecological time series could help distinguish these possibilities [2]. Using permutation entropy as a model-free measure of time-series complexity, they compared estimates of intrinsic predictability with actual forecasting error across simulations and hundreds of empirical ecological time series.

The important conceptual distinction is between intrinsic predictability—how predictable a system could in principle be—and realised predictability—how well a particular forecasting approach actually performs. The gap between the two becomes informative. A large gap suggests scope for improving models or data; a small gap may indicate that the dynamics themselves impose a stronger limit on prediction.

Testing forecastability experimentally

Experimental microbial ecology provides an unusually useful setting in which to test these ideas. Daugaard, Munch, Inauen, Pennekamp and Petchey followed microbial communities for five months under constant and fluctuating temperatures [3]. Rather than simply generating forecasts, they asked which properties of the ecological system determined forecast skill. Species interactions turned out to matter.

Species connected to more other species tended to have individually weaker interactions and were, somewhat counterintuitively, better predicted. Increasing the number of variables included in forecasting models also generally improved forecast skill. Meanwhile, fluctuating environmental conditions reduced forecast skill for some—but not all—species [3].

Ecological complexity therefore did not translate into a simple rule that “more complexity means less predictability”. Instead, forecastability depended on the structure and strength of interactions and on environmental context.

Does biodiversity make ecosystems harder to forecast?

This raises another interesting possibility. If communities containing more species have more interactions and more complicated dynamics, perhaps biodiversity itself makes ecosystems less predictable. Limberger and colleagues recently tested this experimentally using 30 protist communities observed for 41 weeks and 123 sampling dates [4]. Species richness was manipulated, while communities experienced either constant or gradually declining light.

The results provided no simple relationship between biodiversity and forecastability. For individual species, greater richness tended to reduce forecast accuracy under constant light but increase it under declining light. These effects were relatively weak and were partly associated with differences in variability and temporal autocorrelation. Forecasts of aggregate ecosystem properties such as community biomass and oxygen concentration showed little dependence on either richness or the light treatment [4].

Higher diversity, in other words, did not simply make ecological dynamics less forecastable. This is an important empirical result because assumptions about ecological complexity are easily made but difficult to test. Long, highly replicated microcosm experiments make those assumptions experimentally tractable.

Prediction as a scientific method

This work also reflects a broader argument about how ecology can use prediction. Pennekamp, Petchey and colleagues compared ecological prediction with practices in applied fields such as epidemiology and other disciplines where forecasts are routinely generated and subsequently evaluated [5]. They argued for more explicit and consistent evaluation of predictive proficiency in ecology.

Forecasting is valuable here not only because someone might need to know what an ecosystem will look like next month. It also provides a demanding test of ecological understanding. A model can often explain observations from which it was constructed. Predicting observations that have not yet occurred is harder. When forecasts fail, their failures can reveal missing mechanisms, inappropriate assumptions or insufficient observations. The scientific cycle therefore becomes: observe → model → forecast → observe again → evaluate → improve.

Towards a more predictive ecology

Ecological forecasting connects many of the themes running through Petchey and colleagues’ research. Functional diversity asks which differences among organisms matter for ecosystem processes. Video microscopy makes it possible to quantify some of those differences at the level of individuals. Experimental microbial ecology allows communities and environmental conditions to be manipulated. Response diversity asks how variation in species’ environmental responses contributes to stability.

Forecasting asks whether all of that ecological knowledge ultimately improves our ability to predict what happens next. Just as importantly, it asks where the limits lie. The goal is therefore not a claim that ecosystems should be perfectly predictable. Ecological systems contain stochasticity, environmental uncertainty, nonlinear interactions and other sources of unpredictability. Instead, the research programme seeks to make predictability itself quantitative: which ecological variables can we forecast, how far ahead, with what uncertainty, and why does forecast skill differ among systems?

References

[1] Petchey, O. L., Pontarp, M., Massie, T. M., Kéfi, S., Ozgul, A., Weilenmann, M., et al., Pennekamp, F. & Pearse, I. (2015). The ecological forecast horizon, and examples of its uses and determinants. Ecology Letters, 18, 597–611. DOI: 10.1111/ele.12443. Consensus citation count: 304.

[2] Pennekamp, F., Iles, A. C., Garland, J., Brennan, G., Brose, U., Gaedke, U., et al. & Petchey, O. L. (2018). The intrinsic predictability of ecological time series and its potential to guide forecasting. bioRxiv. DOI: 10.1101/350017. Consensus citation count: 116.

[3] Daugaard, U., Munch, S., Inauen, D., Pennekamp, F. & Petchey, O. L. (2022). Forecasting in the face of ecological complexity: Number and strength of species interactions determine forecast skill in ecological communities. Ecology Letters, 25, 1974–1985. DOI: 10.1111/ele.14070. Consensus citation count: 21.

[4] Limberger, R., Daugaard, U., Choffat, Y., Gupta, A., Jelić, M., Jyrkinen, S., et al., Pennekamp, F., et al. & Petchey, O. L. (2025). Mixed Evidence for Species Diversity Affecting Ecological Forecasts in Constant Versus Declining Light. Global Change Biology, 31. DOI: 10.1111/gcb.70364. Consensus citation count: 1.

[5] Pennekamp, F., Adamson, M. W., Petchey, O. L., Poggiale, J., Aguiar, M., Kooi, B., Botkin, D. & DeAngelis, D. (2017). The practice of prediction: What can ecologists learn from applied, ecology-related fields?. Ecological Complexity, 32, 156–167. DOI: 10.1016/j.ecocom.2016.12.005. Consensus citation count: 28.