Skip to content

Detecting predictability in complex ecosystems

A couple people pointed me to a recent article, “Detecting Causality in Complex Ecosystems,” by fisheries researchers George Sugihara, Robert May, Hao Ye, Chih-hao Hsieh, Ethan Deyle, Michael Fogarty, and Stephan Munch.

I don’t know anything about ecology research but I could imagine this method being useful in that field. I can’t see the approach doing much in political science, where I think their stated goal of “identifying causal networks” is typically irrelevant.

That said, if you replace the word “causality” by “predictability” everywhere in the paper, it starts to make a lot more sense. As they write, they are working within “a framework that uses predictability as opposed to correlation to identify causation between time-series variables.” Setting causation aside, predictability is an important topic in itself. The search for patterns of predictability in complex structures may motivate causal hypotheses that can be studied more directly, using more traditional statistical designs such as experiments and observational studies.


  1. Sherman Dorn says:

    The equation look sort of like the basic Lotka-Volterra difference equations, except without the population interactions. Am I wrong in assuming that interaction terms would make the dynamics more (obviously) chaotic?

  2. Dan says:

    I don’t think “fisheries researcher” is the best description of Robert May (or Lord May of Oxford, if you prefer…)

    • Andrew says:

      OK, 6 fisheries researchers and a zoologist. Whatever.

      • peter sunstone says:

        Actually, 5 fisheries researchers. Sugihara was a Managing Director at Deutsche Bank (wikipedia). So a Banker a Baron and 5 real people. I would like to know the causality in that!

        • Ben Bolker says:

          It’s starting to sound like “The Hunting of the Snark” (Wikipedia : “The crew consists of ten members, whose descriptions all begin with the letter B: a Bellman (the leader), a Boots, a Bonnet-maker, a Barrister, a Broker, a Billiard-marker, a Banker, a Butcher, a Baker, and a Beaver.” (No Barons.)

          More seriously: I’ve been aware of this general approach for a while, ever since it was being used for deterministic-chaos hunting in the 1990s. It seems sensible, but I have always been torn — the group has made very strong claims for its capabilities, which I am dubious of on general principles (of “ecology is harder than that”), *but* I have never done my due diligence of digging in and really evaluating it carefully for myself.

  3. Anonymous says:

    Aggregated polls are a good way to predict presidential election outcomes. But this is not very useful knowledge if you are running a presidential contender, and want to know how to intervene in the world to win it.

    Silly example but I think political science is a lot more consequential that the emphasis on prediction suggests. In this view causality matters, a lot.

  4. LJ says:

    I don’t agree with an idea decorating a title giving a fancier name as they used “causal” instead of “predictive” for a time series model. However, there are quite a lot of people who assume a time order as “causality” and it is really hard to disprove them and also counterfactual modeling is not always feasible. Dose anyone who can suggest me a better way to address this?