The Wolf's Invisible Gift: Why Removing Yellowstone's Predators Could Crash Its Bison

In the predator-free decades before wolves were reintroduced to Yellowstone, something strange happened to the bison. Hunters and ecologists had long assumed that removing the great cats and wolves would let prey like bison thrive — that's the whole logic of a trophic cascade, after all. Instead, a new model built from the bones of thousands of real elk, bison, wolves, and cougars suggests the opposite: bison declined when predators vanished, and recovered when they returned. The predators were not just killing bison. They were, indirectly, feeding them.

That counterintuitive result — that losing a predator can be bad for the prey — is one glimpse into a much larger ambition. Willem Bonnaffé and colleagues at Oxford, Utah State, and Yellowstone National Park have built something ecologists have long wanted and never quite achieved: an ecosystem model that starts from the body of a single animal and scales, molecule by molecule, year by year, all the way up to the flow of nutrients through an entire national park. They call it an ecosystem integral projection model, or Eco-IPM. It is the first such model to track the body mass distributions of six structured species — two plants, two herbivores, two carnivores — plus the decomposers, the soil nutrients, and the organic matter that connect them.

The ecological stakes are concrete. Yellowstone is the stage for one of the most-watched wildlife experiments of the modern era: the extirpation of wolves and cougars in the 1920s, and their reintroduction beginning in 1995. The paper's model reproduces the known signatures of that history — and then pushes beyond them to make testable predictions. It suggests wolves and cougars have played a positive, measurable role in the rise of bison in northern Yellowstone, a finding that runs against intuition and demands explanation.

The Science

The challenge here is one of scales. Individual elk graze, grow, reproduce, and die. Populations rise and fall. Ecosystems cycle nutrients. These levels are connected, but the connections are not simple: a predator doesn't just kill prey, it changes prey behavior, which changes what vegetation the prey eats, which changes the soil, which changes the plants, which changes everything again. And it all depends on how big each individual is.

Small trout eat insects; large trout eat other fish. A young elk can be caught by a wolf; a mature bull much less so. A baby aspen is within browsing reach of elk; a tall aspen is not. Body size — the continuous, ever-changing phenotype of every organism — sits at the center of nearly every ecological interaction. But most population models either ignore body size or chop it into a few arbitrary discrete bins (juvenile, adult). Integral projection models (IPMs) avoid that problem by describing how a population's distribution of a continuous trait like body mass changes from one year to the next, through survival, growth, and reproduction. They ask: for every possible starting body mass, how many individuals survive, how big do they get, and how many offspring do they produce?

IPMs have been brilliant at single-population questions. But, as the authors note, they have rarely stretched beyond one or two species and a single trophic level or two. No one had built an IPM that puts population structure into plants, herbivores and carnivores at the same time, with multiple competing species at each level, wired together by a recycling loop of nutrients.

The Eco-IPM does exactly that. Its central equation, a single integrodifference equation, describes how the density of individuals of each species $N^{(i)}(z,t)$ with body mass $z$ evolves year to year:

$$N^{(i)}(z,t+1)={\int }_{S_z}\{f_r(z|z^{\prime })p_b(z^{\prime })n_o(z^{\prime })+f_g(z|z^{\prime })\}{p}_s(z^{\prime })N^{(i)}(z^{\prime },t)d{z}^{\prime }.$$

Read it as a sequence of biological steps. First, an individual of mass $z^{\prime }$ survives with probability $p_s(z^{\prime })$ — which itself is the product of surviving intrinsic causes (disease, senescence) and surviving being eaten. Surviving individuals then either reproduce, producing $n_o(z^{\prime })$ offspring whose size is drawn from a recruitment distribution $f_r$, or grow to a new mass $z$ drawn from a growth distribution $f_g$. The whole thing integrates over all possible starting masses, evolving the entire distribution forward one year.

The crucial ingredient is bioenergetics: the model ties every demographic rate to the actual biomass an individual can capture and convert. An animal of mass $z$ ingests a certain biomass $Z_c$ from its environment — grass for elk, tissue for wolves — assimilates a fraction of it, and then must allocate the result among three demands: metabolic maintenance $Z_m$ (the running cost of being alive, which scales with mass to the ¾ power), reproduction $Z_r$, and growth $Z_g$. If energy is scarce, growth and reproduction suffer. If it's abundant, individuals get bigger, and bigger individuals survive better and reproduce more. This single bookkeeping principle is what lets the model scale from one individual to an ecosystem: the fluxes of biomass between species are literally the fluxes of biomass within individuals.

Around this skeleton, the model adds the ecosystem furniture: decomposers that break down dead organic matter into nutrients, nutrients that plants absorb, and a recycling loop that closes the system. Dead bison and elk become organic matter, which decomposers turn into nutrients, which grow the grass that feeds the living bison and elk. Predators, through scavenging carcasses and killing prey, feed back into that loop. The result is a six-species, three-trophic-level web where every arrow is a kg of biomass per year.

What They Found

The proof of concept is Yellowstone. The authors parameterized the model for two competing woody plants (aspen and willow), two herbivores (elk and bison), and two carnivores (wolves and cougars), plus grass, decomposers, nutrients, and organic matter. Then they fitted it to decades of real count data from northern Yellowstone using a new optimization algorithm — a sparse-noise differential-evolution Monte-Carlo scheme that efficiently searches a high-dimensional parameter space (Bonnaffé et al., 2026).

First, they validated the machinery on artificial data: they simulated the model with literature-derived parameters, perturbed them with noise, and asked whether the fitting algorithm could recover the original behavior.

shows the fit before and after training pulling back toward the true dynamics. Then they applied it to the real time series (

), and the fitted model landed on a stable equilibrium where all six species coexist — a necessary reality check, since the real northern Yellowstone system has persisted for decades.

The headline result comes from a controlled experiment the model makes possible but the real world never will: remove the predators from an ecosystem at equilibrium, wait 200 years, then bring them back.

and

show the consequences. Following predator extirpation, elk numbers rise — the classic trophic cascade. Aspen, released from heavy browsing pressure, regenerate. All as expected.

But bison do the opposite of what a naive cascade predicts. They fall after predator removal and climb after reintroduction. The model attributes this to competition. With wolves and cougars gone, elk — the more abundant competitor — expand freely and outcompete bison for grass. When the predators return, they suppress elk, and bison, released from competition, recover. The model's prediction is unequivocal: wolves and cougars make a positive contribution to bison numbers in northern Yellowstone, not by sparing them but by restraining their rivals (Bonnaffé et al., 2026).

The model also reproduces a known indirect effect: predators promote woody deciduous vegetation regeneration.

A shows aspen biomass distributions responding to the perturbation, and the elasticities in

E–F show where the effects concentrate. This matching of real-world patterns — the rise of aspen after wolf reintroduction documented in Yellowstone — lends the model credibility.

None of this is a simple story. The key facts below distill the three most striking results: the surprise that predator removal hurts bison, the reproduction of the vegetation regeneration effect, and the demonstration that a six-species, three-trophic-, nutrient-cycling model can be fitted to real count data at all.

Why This Changes Things

For decades, the Yellowstone story has been told largely as a linear tale: wolves return, elk decrease, vegetation recovers. The Eco-IPM shows the full web is far stranger. Predators exert their influence not only through whom they kill but through whom they protect by default. By suppressing the dominant herbivore, the predators free up resources for the subordinate one. Call it a "competitive release in reverse": predator removal didn't release bison, it released their rival.

This matters far beyond Yellowstone. Ecology has long had a ladder of models, each rung sacrificing one kind of realism for another. Individual-based models track every agent but become computationally intractable at ecosystem scale. Aggregate ecosystem models track total biomass but blur away the individual variation — the small elk that die and the large elk that don't — that actually drives dynamics. The Eco-IPM occupies a neglected middle rung: it keeps the continuous distribution of body size for every species, preserves the bioenergetic bookkeeping that couples the levels, and still remains, demonstrably, fittable to real data. That combination has enormous practical value.

The fitting algorithm is itself a contribution. High-dimensional ecosystem models have historically been almost impossible to calibrate — the parameter space is too large and rugged. The authors' sparse-noise differential-evolution scheme, validated on simulated data before being trusted on real data, opens the door to grounding large structured models in real ecosystems rather than leaving them as theoretical exercises.

There's a sobering lesson embedded too. Models like this are only as good as the binary assumptions they make. The authors are candid: intrinsic and consumer-caused mortality are treated as independent, which in reality they are not — a sick animal is more vulnerable to a wolf. Diseases and hunting are ignored. Direct competition (physical interference) is set aside in favor of resource competition alone. These simplifications keep the model tractable, but they are precisely the axes along which real systems may deviate from its predictions.

What's Next

The framework is deliberately modular. Adding a new species means supplying its parameter set — survival curves, metabolic costs, diet fractions — not redesigning the machinery. The authors pitch it as suitable for savannahs, marine systems, and beyond, with implications for population management and ecological forecasting. One can imagine using it to ask what happens to an ecosystem if a keystone herbivore is culled, or to plan a reintroduction, or to forecast how a changing climate reshapes the balance between competitors through its effects on growth and survival.

Open questions abound. Does direct competition matter enough to change the qualitative results? Do the bison predictions survive the relaxation of the independence assumptions? The model is a scaffold, not a finished cathedral — but it's the first scaffold that holds up all six species at once.

Perhaps the deepest message is about how ecological insight gets made. The bison result was not obvious from watching either species alone, nor from any single study. It emerged only when the full web — with individual body masses, bioenergetic budgets, and nutrient recycling — was assembled and allowed to run. That is what the model offers: not a single answer to be memorized, but a way of asking questions that a field ecologist or a park manager could never ask in the wild, because the wild doesn't run controlled 200-year experiments. In Bonnaffé and colleagues' model, it does.