A weather icon is the end of a much longer process. Observations, estimates of the current atmosphere, models, interpretation and communication all shape what the reader sees. Understanding that process helps explain why forecasts can improve, change and still remain uncertain.
The gold standard of weather forecasting is a clearly defined prediction whose production, uncertainty and performance can be examined. A precise-looking number is useful only when we know its location, valid time, variable and limitations.
This guide develops the observation-to-prediction branch of Weather. All numerical cases are invented learning exercises, not current forecasts or safety advice. For real conditions, consult the responsible meteorological service and follow current official warnings and instructions.
eduKateSG · Weather Forecasting
Find your next learning step
Follow the forecasting chain from observations to a defined prediction, an uncertainty statement and an appropriate check.
Full chapter index · Practice and transfer · Education root · How X Works
Full chapter index
Define the forecast · Chapters 1–2
Build the prediction · Chapters 3–4
Interpret uncertainty · Chapters 5–7
Verify and transfer · Chapters 8–10
Follow the specialists · Chapters 11–12
Write down what is being predicted, where, when and on what basis. A temperature at one station, a regional rainfall accumulation and a warning about possible impacts are different products. They should not be judged using an undefined idea of “correct weather”.
An invented forecast might concern at least 1 mm of rain at a specified location between 14:00 and 18:00 local time tomorrow. Its event is not “some rain somewhere during the day”. Changing the place, threshold or interval changes the question.
Also separate issue time from valid time. A forecast issued at 08:00 for 16:00 has an eight-hour lead time. A revised forecast issued at 14:00 for the same valid time has only two hours of lead time. Comparing them without acknowledging that difference can mislead.
Observations describe measured conditions at particular places and times. Instruments, coverage and quality control matter. A reading from one station is not automatically representative of every nearby street or altitude.
A weather map can combine direct measurements with estimates and model products. Read the product documentation rather than assuming every coloured grid cell contains its own physical sensor. Likewise, radar reflectivity is not simply a direct rain-gauge reading for each point.
For the measurement branch, use Measurement and Weather Radar and Rainfall Estimation. Begin with what the instrument actually measures, then trace the reasoning that produces the variable a reader wants.
CHAPTER 3 OF 12 · Build the prediction
3. Data assimilation builds a starting estimate
Numerical prediction requires an estimate of the atmosphere’s initial state. Data assimilation combines observations with information from a prior forecast or background state, using assumptions about their uncertainties. It is not just pasting the newest measurements onto an otherwise complete map.
The Met Office data-assimilation reference describes methods for this task. The starting state is an analysed estimate, not perfect knowledge of every atmospheric quantity everywhere.
In a learning review, ask what observations were available at the forecast’s initialization time. Information collected later may help verification, but it was not necessarily available to the original forecaster. Keeping that distinction prevents an unfair comparison between a real forecast and a hindsight reconstruction.
A numerical weather model represents atmospheric processes using equations and computational methods. Finite resolution and approximations mean that a model grid is not a miniature atmosphere containing every real feature.
Resolution describes one aspect of the representation; it does not independently guarantee accuracy. Observational quality, initial-state estimates, model formulation and the situation being predicted also matter. Avoid deciding which forecast is best solely from the smallest advertised grid spacing.
The Met Office numerical-prediction overview provides a concrete example of model systems. Connect this with Weather Forecasting, Differential Equations and Numerical Models. A useful explanation names the represented process and the scale at which the output is being interpreted.
An ensemble produces multiple forecasts with variations designed to represent relevant uncertainties. Comparing members helps investigate possible outcomes rather than treating one trajectory as the only plausible future.
The Met Office ensemble research overview distinguishes work on uncertainties in initial conditions and model representation. Ensemble members are related model outputs, not independent eyewitnesses who have visited tomorrow.
In an invented fifty-member ensemble, twenty members exceed a stated rainfall threshold at the stated place and time. The member fraction is 20/50 = 0.40. It is a model-based estimate to interpret and verify, not proof that nature’s probability is known exactly to two decimal places. Check how the provider converts model output into the published product.
A rain probability needs its provider’s definition. It should not automatically be interpreted as the percentage of the day that will be wet, the percentage of an area that will receive rain or the severity of the rain.
Consider two teaching events: any measurable rain during a six-hour period, and more than 20 mm during one hour. They are not interchangeable. A high probability for the first does not specify the probability or consequences of the second.
The Met Office discussion of ensemble use connects member proportions with event probability. For a learner, the first repair is often not more arithmetic but a better event definition. Write the threshold, location and interval beside the percentage before drawing a conclusion.
Separate hourly probabilities do not determine a whole-window occurrence probability without their overlap. Combining Rain-Occurrence Probabilities Across Time Windows builds original complete rainfall scenarios, derives dependence-free bounds and distinguishes marginal chances, conditional hazards and accumulated thresholds.
A changed forecast may reflect new observations, a new initialization, a different model cycle or interpretation of uncertainty. An update is not, by itself, proof that the earlier forecast was irresponsible or that the later one is certain.
Record the issue time and compare like products. A weather forecast, a nowcast and an impact warning can serve different purposes. Official warnings and the instructions attached to them should not be replaced by a classroom probability rule.
The Met Office discussion of exact-date headlines explains why apparent certainty can exceed the evidence. For communication, say what changed and what uncertainty remains. Do not remove an important limitation merely to fit an attractive headline.
Forecast verification compares predictions with a suitable reference for the stated variable, place and valid time. Choice of sample and reference matters. A forecast for a region should not be casually treated as a prediction for one unrepresentative instrument.
| Part of the record | Question to ask | Why it matters |
|---|---|---|
| Forecast issue time | When was it available? | Defines lead time and available information |
| Event definition | Which threshold, location and period? | Defines what counts as an outcome |
| Reference observation | How was the outcome determined? | Controls the comparison basis |
| Verification sample | Which cases were included? | Limits generalisation |
| Baseline | What simple alternative is used? | Distinguishes error from skill |
The NOAA Environmental Modeling Center verification resources show verification as part of model assessment and development. One memorable failure or success is not a complete performance record.
Comparing Displaced Rainfall Patterns with Neighbourhood Fractions reconstructs a displaced rainfall pattern at several neighbourhood widths, then checks dry cases, masks, location limits and the distinction between pooling score components and averaging scores.
Take three invented forecasts of 29, 31 and 30°C, with corresponding observations of 30, 29 and 31°C. Define signed error as forecast minus observation. The errors are −1, +2 and −1°C.
The mean signed error is zero, because the positive and negative errors cancel. The mean absolute error is (1 + 2 + 1)/3 = 4/3, approximately 1.33°C. Zero average bias therefore does not mean every forecast was correct.
Keep the scope visible: this is three cases under one invented comparison. It does not establish seasonal accuracy or superiority over another service. To claim skill, choose a relevant baseline and compare on the same cases and lead times. A score without that context can encourage a much stronger claim than the arithmetic supports.
Task A: Thirty of fifty ensemble members exceed a defined threshold. Find the member fraction. Check: 30/50 = 0.60. What cannot be inferred? The exact timing within the interval, the amount above the threshold, and a guarantee that the event will occur.
Task B: A forecast issued at 06:00 is valid at 18:00 the same day; a second is issued at 15:00 for that same valid time. Find the lead times. Check: twelve hours and three hours. Explain why equal errors would not necessarily mean equal predictive performance.
Transfer task: Change from temperature to rainfall. The habit of defining the variable and reference still applies, but the appropriate event, spatial comparison and error measure may change. Explain what should be retained and what needs redesign before reusing the temperature evaluation.
Forecast production and interpretation
- Weather — retain the wider atmospheric context.
- Forecasting — compare prediction across domains.
- Measurement — inspect evidence quality.
- Data Literacy — read uncertainty and scope.
- Numerical Weather Models — connect equations with prediction.
- Radar and Rainfall Estimation — distinguish instruments and derived quantities.
Performance and wider systems
- Forecast Reliability, Sharpness and Decision Thresholds — develop probabilistic verification.
- How Forecasting Works — follow evidence into expectations.
- Flood Early Warning — distinguish hazard information and response systems.
- Energy Forecasting — explore a weather-dependent application.
- Singapore Weather Radar and Climate Services — connect capabilities with a local system.
Individually useful forecast decisions can compete for one limited resource. Allocating Limited Protection Across Simultaneous Forecast Decisions compares complete feasible plans in original binary allocation examples, tests ranking shortcuts and keeps probability verification separate from the resource and consequence model.
Does a 60% probability guarantee rain?
No. Interpret the provider’s event definition and evaluate probabilities across an appropriate collection of comparable cases, not as a promise for one case.
Does a more detailed map guarantee accuracy?
No. Display detail and model resolution do not alone establish forecast performance.
What should I practise first?
Take an archived forecast, identify its issue and valid times and state the exact predicted quantity. Only then choose suitable observations and a comparison method.
- Met Office: Data Assimilation — the initial-state problem.
- Met Office: Numerical Prediction — model representation.
- Met Office: Ensemble Forecasting — uncertainty.
- NOAA: Forecast Verification — evaluation resources.
Continue through Education and How X Works. A useful forecast explanation keeps the observation, model, uncertainty and check connected.