Over the past 50 years, computational power, theory, and data collection have improved, leading to a revolution in weather forecasting. What began as an inferior practice quickly overtook other methods. Computational predictions have now replaced traditional subjective predictions, and detailed weather forecasts are integral parts of our everyday lives. The figure below (adapted from Shuman 1989) shows how computational weather forecasting models have progressed since the 1950s.
Figure adapted from Shuman 1989. Error in numerical weather forecasts has decreased substantially since the 1950s due to improvements in the models and the computers.
In contrast to weather forecasts, there are very few ecosystem forecasts available. Those that are available are at the cutting edge of science, and are often released only after the passage of the events they are trying to predict. Yet I would argue that we could be on the cusp of a similar revolution in ecological forecasting.
Many lessons from the past 50 years of weather prediction can be carried over to the development of systems designed for ecological forecasting. There are three key components to a good forecasting system, all of which we can begin to implement and/or take advantage of right away. These are: monitoring (steady streams of data), adaptive computational algorithms, and forecast availability. The common service provided by all of these components, and the crux, I would argue, to a revolution in ecosystem forecasting, is feedback.
I've sketched out below a concept map of an ecological forecasting system, highlighting the mechanisms of feedback. Briefly, data is input to the algorithm, which produces a forecast. This forecast then becomes available to a number of feedback avenues, including the designer, users, and the algorithm itself. (Side note: I used the "Concept Map Builder" designed by the Center for Ocean Sciences Education Excellence, which is a tool under development, worth checking out.)
Monitoring is, of course, a critical component as it provides input data. In this system, however, data plays another role. Suppose we are producing weekly forecasts. Then each week, we want to know how well we did with our forecast. At first, our forecasts might be little better than a random guess. That's okay. As long as we have the data coming in to tell us how well we're doing. When we're not doing well, we want to know so that we can adjust.
That brings us to the second component: adaptive computational algorithms. Using techniques borrowed from computer sciences, such as the genetic algorithm, a system can learn based on its past successes and failures. It's important for the system to be able to assimilate a steady stream of data and to adapt based on that. Ideally, a system should be flexible enough to incorporate new theoretical information as well.
The third component, forecast availability, is one that has been missing. In my opinion, this has been holding back progress in ecosystem forecasting. I suggest that, like numerical weather forecasters, we should be putting our forecasts out there, even while they are little better than subjective forecasts. It's okay if many forecasts fail. With an ensemble of forecasts available, and with a steady stream of observations coming in, our computational systems can begin to evaluate, learn, and adapt. Over time, the failures will presumably be fewer and fewer, and the forecasts will be more and more useful.
There are a number of subtleties that I've brushed over here, for the sake of simplicity and generality, as well as some practical considerations that are nontrivial. Nevertheless, these are the basic ingredients, as I see it, to a revolution in ecosystem forecasting. I'll be presenting these ideas at the Ocean Sciences meeting in Portland, Oregon next week. If this sort of thing is your bag, look me up.