Ecology Letters 2022 PhD chapter 1

Forecasting in the face of ecological complexity

In a small microbial community, species with many but weak interactions are easier to forecast than species with few strong ones. How connected a species is turns out to be a forecasting trait in its own right.

Daugaard U., Munch S., Inauen D., Pennekamp F. & Petchey O.L.

doi: 10.1111/ele.14070 ↗

01 · The question

Is ecological forecasting feasible given community complexity?

Ecology aims to forecast how populations and communities respond to change, and complexity is usually cast as the obstacle. Across systems, more species and more links are expected to mean less forecast skill. Within one system, though, species differ in how many partners they interact with and how strongly, and which end of that trade-off is easier to predict had not been tested.

Two-panel schematic. Left: across systems, complexity and number of links rise while understanding and forecast skill fall. Right: within one system, species trade many interactions against strong ones, and the consequence for forecast skill is left open. — Fig. 1 Two senses of complexity. Across systems, more links are expected to erode forecast skill. Within one system, the direction was an open question.

02 · What we did

A microbial community, five months, two thermal regimes

We grew a tri-trophic aquatic microbial community of bacteria, algae, ciliates, flagellates and a rotifer in eighteen bottles for 154 days. Nine sat at a constant 17.3 °C; the other nine were split across three fluctuating series with the same mean, variance and autocorrelation, one of them copied from a nearby stream. We sampled three times a week, giving 66 abundance points per bottle, measured by flow cytometry, imaging or manual counts depending on how big the organism was.

Left: the microbial food web annotated by functional group, from three grouped bacteria and an autotroph up to a top predator. Right: the constant and three fluctuating temperature series, and the eighteen experimental bottles. — Fig. 2 The community and the design. Left: the food web, from grouped bacteria up to a single top predator. Right: one constant and three fluctuating temperature series, and the eighteen bottles spread across them.

For each of eight target species we forecast abundance with multiview embedding, an empirical dynamic modelling method: train on the first 44 time points, predict the last 22, score the error. Separately, in an analysis of its own rather than as a by-product of the forecasts, we used convergent cross-mapping to find which variables influenced each target, and S-map Jacobians to estimate how strongly they did so. Those two quantities, the number of interactions and their mean strength, are what we mean by a species’ connectedness.

03 · What we found

Generalists are easier to forecast than specialists

Forecast error fell as a species’ number of interactions rose, and climbed as their mean strength rose. The two run against each other: the more interactions a species had, the weaker they were. Generalists, with many weak links, were therefore the easiest to forecast and specialists the hardest. Neither relation depended on the temperature regime.

Three scatter panels, each split into constant and fluctuating temperature: forecast error falls as the number of interactions rises, rises with mean interaction strength, and mean strength itself falls as the number of interactions rises. — Fig. 3 Forecast error (RMSE, lower is better) falls as a species' number of interactions rises (A) and climbs with their mean strength (B); species with more interactions interact more weakly (C). Constant (circles) and fluctuating (diamonds) regimes give the same picture.

Connectedness is a predictability signature. Where a species sits in the community accounts for about a third of the variation in forecast error: a strong tendency, not a hard limit.

Model size mattered too: the more state variables a model used as predictors, the lower its error, in all eight species. Fluctuating temperatures raised forecast error for three of the eight (C. reinhardtii, P. caudatum and the rotifer) but left the interaction patterns themselves untouched, which argues against any universal law tying complexity to predictability.

Eight per-species panels of median forecast error against the log number of predictors used, overlaid for constant and fluctuating temperature, with three species boxed in red. — Fig. 4 Median forecast error against the number of predictors used, species by species. Error falls as models grow. Fluctuating temperatures cost skill in the three boxed species and left the other five alone.

04 · Why it matters

Predictability is structured by the community itself

A species’ interactions are a forecasting trait. Estimating connectedness says something about which parts of a community will be easy or hard to predict before any forecasting is done, and so about where monitoring effort and model attention are worth spending.