 So, one of the questions that really interests me about ecological forecasting is to thinking about as a discipline in ecology, can we make ecology more predictive and kind of thinking about can we forecast ecology in a way similar to the way we currently forecast weather? One of the things that's important in answering that question is having some measure of predictability to judge whether we can predict ecological systems or not. There are multiple possible ways to judge predictability, but one of the ones that I found very useful is to think about the uncertainty in our predictions and how those grow through time. If you've made a graph looking at the uncertainty in your prediction versus time it should start at some level, you know, what is our understanding of the system right this moment and as I move into the future that uncertainty should increase in time. In general, as uncertainties compound we're always going to see the pattern that as we move out into the future our uncertainty about the state of the system is going to grow. So one measure of predictability would be what is the rate at which that uncertainty is growing and another might be when does that uncertainty get to some point where it's essentially indistinguishable from our kind of background understanding. So it might converge to kind of the average understanding of the system and sometimes we're not doing any better than than random just resampling the data we already have. We just are kind of converging to the range of natural variability in a system. So how do we understand predictability and what affects it? One of the things that we can do is think about the classes of models that we use to make predictions in ecology and one very common class of models are what we call dynamic models. Models that try to predict the future of the system as a function of the current state of the system and then possibly some other sets of covariates and those models would then have some sets of parameters and then some process error that captures the the factors that we don't understand about the system. What we can then do from that general understanding of dynamic models is we can actually say well how does the variance in a dynamic model increase through time? What we can do is partition that uncertainty in our forecast into five key terms and these correspond to the key uncertainties that go into the model and then think about how those uncertainties that go in translate into the predictions. So first we have the uncertainty about the initial conditions. So what is our uncertainty about the current state of the system? We have uncertainties related to the inputs so the drivers so this is the uncertainty about how the future environment is going to change and the sensitivity of our system to those environmental drivers. We can then think about the parameters in our model. We can actually break that down into two components. We can think about what is the uncertainty about the mean value of those parameters and we can also think about how does those parameters vary. So one of the common characteristics of ecological systems is there's often unexplained heterogeneity and unexplained variabilities. We might have year-to-year variability that we don't fully understand or site-to-site variability we'd understand or variability among individuals that we can quantify but we don't yet know the explanatory variables that describe why we see that variability and then finally there's this process error the unexplained components. If we dive into each of these terms we can see that they follow a very common and repeating pattern which is the contribution of each uncertainty the overall predictive uncertainty can be thought of as the product of the uncertainty itself of the input and then how the system responds to that uncertainty. Mathematically that would be expressed as you know the variance of one of these five inputs and then the the sensitivity which is essentially the slope of the relationship between that input and the output we're trying to predict. So mathematically that would be expressed in terms of the derivative which is again a measure of sensitivity. This tells us at a high level that if we want to understand systems and the predictability systems we need to understand the uncertainties about the inputs and we need to understand these sensitivities. One important problem facing ecological forecasters is determining the relative importance of these five different uncertainties for different classes of problems we're trying to make forecasts for. So to give an example the first term that we talk about is the uncertainty about the initial conditions and the sensitivity of systems to their initial conditions. That sensitivity to initial conditions is what ecologists often think of as the stability of the system and at a high level it might tell us whether a system has stabilizing feedbacks that cause it to go to some sort of equilibrium or whether it does not have stabilizing feedbacks but instead is chaotic. One important example of a chaotic system when it comes to forecasting is the atmosphere. So back in the 50s and 60s when numerical weather forecasters were first starting to make predictions they discovered that their models were chaotic. That discovery which is you know innate to the physical equations that describe the atmosphere led them to realize that if they wanted to make predictions that that specific term the uncertainty about the initial conditions was going to dominate over the other uncertainty terms in any prediction they wanted to make over non-trivial time scales. Because of that understanding that that specific terms the internal stability and initial condition uncertainty was the dominant uncertainty over the 50 or 60 years that weather forecasters have been making predictions pretty much the entirety of that approach the entirety of the workflow and the system is largely optimized around that understanding of which uncertainty dominates their prediction problems. So we have billions of dollars in weather satellites and observational campaigns and airplane measurements and buoys and ground measurements and radios on so all these observations that are made on a continuous basis on the atmosphere yes the data is interesting unto itself but one of the main reasons that data is collected is because atmospheric scientists need to constrain the initial conditions of weather forecasts and they need to do this on a continual basis so right now weather forecasts are updated every six hours every six hours they need new information about this the current state of the system to keep the weather forecasts uncertainties from blowing up due to the chaotic nature of their system. By analogy ecological forecasters do not know from first principles which of these five terms are going to dominate the predictions that we're trying to make understanding that is actually an empirical problem it's one that's going to require us to attempt to forecast a large range of different ecological problems and then understand and formally partition out the different uncertainties that contribute to those forecasts. One of things I hope to learn by making ecological forecasts is if there are common patterns to the predictability of different ecological systems so the null model here is that there is no pattern that every time we encounter a new ecological forecast problem it's going to be unique and different than any other ecological forecasting problem we've ever seen. I don't think that's what we're actually going to see I expect that for certain classes of problem we'll see that they're dominated by different sources of uncertainty so there may be ecological forecast problems dominated by sensitivity to the environment there may be ecological forecast problems dominated by you know this chaotic sensitive initial conditions there may be chronically data limited ecological forecast problems that are dominated by the uncertainties in the parameters or there may be certain classes of ecological problems that are dominated by the inherent heterogeneity and variability in ecological systems. I feel that once we have some understanding of that variability in what terms dominate different ecological forecasts it's going to have a real impact on our field. I think it tells us something broadly in a theoretical concept about how ecological systems work about what sorts of processes drive them and I think it does so in a way that really excites me because it allows us to really take a broad comparative approach it allows us to learn things say about the predictability of harmful algal blooms and translate that understanding to trying to predict you know disease dynamics or at least it gives us a common language that ecologists can use across all sub-disciplines to talk to each other about predictability. So in addition to providing us with some theoretical understanding of what drives ecological systems and the potential for seeking generality in ecology and understanding different processes across space is a very practical side to trying to understand which uncertainty dominates which is that if we want to make new predictions for new problems in new systems if we can classify them as being similar to other forecasting problems we know that tells us something about how we should approach that problem. So if we have you know a new emerging invasive species or a new emerging infectious disease and we can say well that is going to be like these other infectious disease or invasive species problems that were dominated by certain types of uncertainty it helps us really focus in on what we should be measuring and how we should be approaching that problem because fundamentally we can't measure everything knowing something about where we should focus our efforts can really make better use and more efficient use of the limited resources that we have. The final reason I think understanding which uncertainties dominate ecological forecasts is important because I think it has a real impact on the methods we use so what we measure should be impacted by which uncertainties dominate how we build our models are going to be reflected in understanding what uncertainties dominate and then how we build statistical tools for bringing models and data together needs to be optimized around what sorts of uncertainties dominate ecological forecasts. One thing we see right now is that there's a lot of experience and tools that we have the potential to borrow from other fields like weather forecasts for developing ecological forecasts however weather forecasts are fundamentally trying to tackle a different uncertainty than ecological forecast that I think there's real value in stepping back and asking well are are those tools and their assumptions really the appropriate tools for ecological forecasts or maybe we need to revisit where those tools were derived from and make slightly different assumptions that are optimized for the types of uncertainties that dominate ecological forecasts.