1. Quick Summary
Forecasting is an initial-value problem: given the current state of the atmosphere, compute how it evolves. Both the observation and the model introduce error.
The atmosphere is chaotic, so small errors in the initial state grow. That growth sets the practical limit on how far ahead specific conditions can be predicted.
Improvements come from better observations, better models and more computing power, and they extend useful range gradually rather than removing the limit.
2. What It Means
Numerical weather prediction solves equations of fluid motion and thermodynamics on a grid. Grid spacing determines which processes are resolved and which must be approximated.
Parameterisation represents processes smaller than the grid — clouds, turbulence, surface exchange — and it is a major source of model error.
Ensemble forecasting runs many simulations from slightly different starting states. The spread between them estimates uncertainty, and a tight ensemble usually means higher confidence.
Verification compares forecasts against observations over time. Skill is measured against a reference, such as climatology or a simpler model, so improvement is defined relative to a baseline.
3. Why It Happens
Chaos amplifies error at a finite rate. Errors roughly double over a characteristic time, so beyond roughly two weeks the forecast carries little more information than climatology.
Observations are uneven. Oceans, the upper atmosphere and parts of the Southern Hemisphere are sampled far less densely than populated land regions.
Resolution costs computing power, and power is finite. Finer grids improve detail but require more computation, so operational centres trade coverage and ensemble size against resolution.
Model error and initial error interact. Improving only one gives diminishing returns, which is why progress requires both better data assimilation and better physics.
Predictability varies by situation. Large-scale patterns can be more predictable than local details, and some regimes — blocking patterns, for example — are systematically harder.
4. Real Examples
Forecasts for a few days ahead are now substantially better than those of several decades ago, and the improvement is largest for large-scale fields such as pressure.
Ensemble systems routinely show a situation days ahead while the deterministic run changes between successive cycles, which is normal behaviour rather than inconsistency.
Tropical cyclone track forecasts have improved faster than intensity forecasts, because track depends more on the large-scale flow that models represent well.
Sub-seasonal and seasonal outlooks describe probabilities rather than specific days, which reflects the predictability limit honestly rather than avoiding it.
5. How It Affects Us
Warnings and decisions depend on communicating uncertainty. An ensemble that shows a range of outcomes is more useful to emergency managers than a single deterministic track.
Users often interpret a changing forecast as unreliability, when a shift within the previously stated range is exactly what a probabilistic system should do.
Economic value is concentrated in extreme events, where accurate warnings prevent damage, which is why verification increasingly weights high-impact cases.
Different sectors need different lead times: aviation hours, agriculture weeks, energy planning months — and each sits at a different distance from the predictability horizon.
6. Key Takeaways
- Skill has genuinely improved and continues to, but chaos sets a firm limit of roughly two weeks.
- Uncertainty is best expressed as an ensemble spread, not as a single forecast.
- Better observations matter as much as better models.
- A forecast that changes within its stated uncertainty range is behaving correctly.