Forecast horizon
A Socratic walk-through of the forecast horizon — reasoned out one step at a time, not lectured.
The question we started with
THE QUESTION #Why can we forecast tomorrow's weather well and next month's not at all?
A three-day forecast today is about as good as a one-day forecast was decades ago. That is a genuine, hard-won improvement, and it invites an obvious extrapolation: keep going and eventually we forecast the month. Yet forecasters will tell you flatly that beyond roughly two weeks there is nothing to be had. Why should progress that has been steady for decades stop at a wall — and is the wall in our computers, or in the sky?
Reasoning it through
REASONING #Begin with what a forecast actually is. You measure the atmosphere's present state as well as you can, write down the equations governing air, and march them forward. The equations are not in doubt, so the only place error can enter is the measurement — and measurement is never exact. There will always be a gap between the state you fed the model and the state the atmosphere is really in.
So ask: what does that small gap do as the calculation runs? It could shrink, stay put, or grow. Edward Lorenz stumbled onto the answer in 1961 while restarting a simulation from a printed number rounded to three decimals instead of the six the machine held. The rerun tracked the original for a while and then diverged completely. Not a bug — a property of the equations.
Why should small differences grow? Because the atmosphere is full of instabilities that feed on gradients, so two nearly identical states can put a developing wave on marginally different sides of a threshold and the difference amplifies. While the error is small it grows roughly exponentially, which means the useful quantity is a doubling time — historically estimated at a couple of days for large weather systems and rather less for smaller ones.
Now the arithmetic that decides the whole argument. If errors double every day or two, what does a better observing system buy you? Halving the initial error buys exactly one doubling time of extra lead. Cut the error by a thousandfold — an almost unimaginable improvement — and you gain about ten doublings, a couple of weeks at the most generous. The pay-off is logarithmic in effort, which is why the limit is not a computing-power problem: no plausible increase in observation or resolution moves it much.
Worse, error does not only grow in place; it cascades upward in scale. Uncertainty in convection you never resolved at all — an individual thunderstorm — contaminates the scale above it within hours, and that contaminates the scale above that. The model cannot escape by getting finer, because refining it merely admits smaller uncertainties that climb the same ladder.
If we cannot remove the uncertainty, can we at least measure it? That is what ensemble forecasting does. Run the model many times from slightly different initial states and instead of one trajectory you get a spread. Where the members stay together, the atmosphere is in a predictable mood; where they fan out by day four, it is not. The forecast becomes a probability rather than a claim, and the amount of predictability available is itself variable — the ensemble is how you find out today's value. That manages chaos. It does not defeat it.
So the horizon of roughly two weeks is a property of the atmosphere, not of the meteorologists. Recent work estimates the intrinsic limit at about two weeks and puts current operational models within a handful of days of that ceiling, so the remaining room for improvement is real but bounded.
Then the puzzle that makes the whole thing click. If we cannot say what the weather does in six weeks, how can anyone say anything about the climate in sixty years?
It is a different question. Forecasting tomorrow is an initial-value problem: given exactly where the system is now, where does this trajectory go? Projecting climate is a boundary-value problem: given the constraints — incoming sunlight, greenhouse gas concentrations, ocean heat capacity, the geography of land and sea — what statistics does the behaviour settle into? The first asks for a point on a path, the second for the shape of the region the path wanders in. Losing track of the point tells you nothing about the shape.
The analogy
THE ANALOGY #Watch a pan of water on a steady flame. You cannot predict where the next bubble will break the surface, and you certainly cannot predict it a minute out — a trace of turbulence decides it. Yet you can say with total confidence that the pan will be boiling in five minutes, and that turning up the flame will make it boil harder. The bubble is weather; the boil is climate.
The pan reaches a steady state while the atmosphere never does — its statistics themselves drift as the boundary conditions change; and the next bubble is unpredictable immediately, whereas the atmosphere is genuinely and usefully predictable for days before it stops being so.
Clarifying the model
THE MODEL #Two refinements matter here.
First, "chaotic" does not mean random or unbounded. A chaotic system is fully deterministic and stays within a bounded set of behaviour; it merely fails to preserve information about which state within that set it started in. That is precisely why the statistics can be stable while the trajectory is not.
Second, the two-week figure is a rough characteristic scale, not a cliff. Predictability is flow-dependent: some situations — a locked-in blocking pattern, a strongly organised tropical oscillation — carry skill further, and some collapse in three days. And skill beyond two weeks does exist for the right quantities, because slowly varying components at the boundary, chiefly ocean surface temperatures and soil moisture, shift the odds of a warm or wet month without pretending to say which days. That is the honest content of a seasonal outlook: a nudge to a probability distribution, not a forecast.
A picture of it
THE PICTURE #How to readRead left to right for how much a prediction depends on knowing the present state exactly, and bottom to top for how tightly external constraints pin the answer down. Tomorrow's weather sits right and low — almost purely an initial-value problem, and the same prediction slides further right as lead time grows, into the corner where nothing survives. Climate in 2100 sits at the opposite extreme because it asks about constraints rather than a trajectory; the tide table is there for calibration, predictable years ahead for exactly the same reason. The seasonal outlook sits between them, drawing what modest skill it has from slow boundary components such as ocean temperature.
What became clearer
WHAT CLEARED #The forecast horizon is not a limit on our machines but a property of the atmosphere: errors we can never eliminate double on a timescale of days, so improving the observations buys only logarithmic gains in lead time. Ensembles convert that limit into an honest probability instead of removing it. And the same chaos that ends weather prediction at two weeks leaves climate projection untouched, because weather asks where a particular trajectory goes and climate asks what shape the set of trajectories has — an initial-value question and a boundary-value question, which happen to share equations but not their answerability.
Where to go next
ONWARD #- How ensemble spread is turned into calibrated probabilities, and how those are scored.
- Why some slow components of the system — ocean temperatures, soil moisture, sea ice — extend skill where the atmosphere alone cannot.
Key terms
TERMS #| Term | What it means |
|---|---|
| Sensitive dependence on initial conditions | the property that arbitrarily small differences in starting state grow into large differences in outcome. |
| Initial-value problem | a prediction whose answer depends on knowing the system's present state; weather forecasting. |
| Boundary-value problem | a prediction whose answer depends on the constraints acting on the system; climate projection. |
| Ensemble forecast | many model runs from perturbed initial states, whose spread estimates the day's available predictability. |
| Attractor | the bounded set of states a dissipative dynamical system settles into wandering around. |
Every term the collection defines is gathered in the glossary.