Did you know that an afternoon weather reading cannot by itself prove a forecast for lunchtime was wrong? The sky changes. A forecast also refers to a place, a period and a weather variable. Before declaring success or failure, a learner must first check that the prediction and the observation answer the same question.
A forecast–observation comparison matches a prediction made in advance with a measured or recorded outcome for the same variable, location and valid time or period. Preserve the forecast as originally issued, record how the observation was collected, compare like with like and describe the remaining uncertainty. That is the immediate answer to ‘How accurate was the forecast?’
The four fields that must match
- Variable: Is the forecast about air temperature, rainfall amount, rain occurrence, wind speed or something else? Different variables cannot be scored as though they were interchangeable.
- Place: Is it a school garden, one weather station, a named district or a large forecast region? A local observation might not represent the entire forecast area.
- Time: Is the forecast for noon, the afternoon or a two-hour window? Note both the issue time and the valid time.
- Definition: Does ‘maximum temperature’ mean the highest reading over a period rather than the reading at one instant? Are rainfall units millimetres or millimetres per hour?
A forecast’s issue time tells us when the prediction was made; its valid time tells us when or for what interval it applies. This distinction matters because a forecast updated at 2 p.m. should not quietly replace the one students recorded at 8 a.m. when they are investigating the accuracy of the morning prediction.
Worked example: a made-up temperature forecast for a class project
Assume the class records the following illustrative, invented values for the same shaded location. On Monday at 8:00 a.m., a hypothetical forecast predicts 30°C at noon. A properly positioned thermometer records 32°C at noon. At 3:00 p.m., the same thermometer reads 29°C. These are made-up teaching values, not data from a Singapore weather station or a claim about an official forecast.
The valid comparison is noon predicted 30°C against noon observed 32°C. Using observed minus forecast, the signed difference is 32°C − 30°C = +2°C. The absolute error is 2°C. The positive sign means the measured noon temperature was higher than the predicted noon temperature in this example. Write down your sign convention, since ‘forecast minus observed’ would give −2°C instead.
Why not compare noon’s 30°C prediction with 3 p.m.’s 29°C reading and report an error of 1°C? Because those values describe different times. A numerically smaller difference obtained from mismatched records is not a better verification. Even a perfectly calculated subtraction cannot repair an invalid pairing.
If the forecast instead said ‘daily maximum 30°C’, the correct comparison would involve the maximum observed over the specified day or valid period, using suitable recorded observations. The noon reading alone would not establish that maximum. The definition of the variable changes the task.
A short diagnostic that catches the misconception
Give learners three cards: ‘Issued 8 a.m.: 30°C for 12 p.m.’, ‘Observed 12 p.m.: 32°C’, and ‘Observed 3 p.m.: 29°C’. Ask which two cards belong in the error calculation, and why. If the student chooses 30 and 29 because the numbers are closer, they are optimising the answer rather than matching the evidence. If they select 30 and 32 but cannot identify the relevant time, return to the idea of valid-time labels.
How to collect and check a small observation log
- First, preserve the forecast: Copy the issue timestamp, forecast location, variable and valid time, as well as the prediction itself. Retain a screenshot or source record when permitted.
- Next, record an observation: Note time, units, place, measurement method and relevant conditions. Do not substitute a different station without labelling the change.
- Then, inspect comparability: Match temperature with temperature and point-in-time with point-in-time. For a period forecast, define an appropriate period outcome.
- Calculate carefully: Write ‘observed minus predicted’ and show the arithmetic and units. Do not turn a 2°C difference into a percentage unless a suitable denominator and reason exist.
- Interpret modestly: One case tells you about one comparison, not a forecast system’s long-term performance. A larger sample with a defined scoring rule is needed for broader claims.
Singapore’s official NEA two-hour forecast dataset describes time-limited forecasts for particular areas. The Meteorological Service Singapore’s current temperature observations explain that automated instrument readings can have gaps or later corrections, and local siting can influence readings. These are useful reasons to keep dataset timestamps and location labels, not reasons to invent missing values.
Progressive practice from guided to independent
- Guided: Use the fictional noon example and write the matched pair together. Have the learner explain why the 3 p.m. number is excluded.
- Supported: Provide four timestamped forecast and measurement cards, with two distractors at different locations. Let the learner build one valid matched pair.
- Independent: Ask for a five-column table: issue time, valid time, place, prediction and observation. Give a new invented dataset and ask for a justified error.
- Transfer: Repeat with an interval prediction such as ‘rain during 1–3 p.m.’. Ask why a dry observation at exactly 1:10 p.m. cannot alone establish whether rain occurred anywhere during the whole interval.
Mistakes worth recognising
A common error is treating a forecast of ‘showers somewhere in an area during a period’ as a promise of rain at every point and minute. Another is updating an original prediction after seeing the result. Both change the question. A third is recording a school thermometer in direct sunlight and comparing it casually with an official station’s shaded-air-temperature reading; measurement method matters.
A practical progress measure is whether students can reject mismatched evidence before calculating. Across three fresh examples, record whether they identify variable, place, time and definition, then give a short explanation of uncertainty. Increasing arithmetic speed is less important than choosing a valid pair.
Questions families and teachers often ask
What if the forecast changed during the day? Preserve each issuance separately. To evaluate the 8 a.m. forecast, do not replace it with an updated noon forecast.
What if one observation is missing? Record it as missing. Do not fill in a convenient temperature or presume that missing means zero.
Does one inaccurate forecast mean forecasting is useless? No. A single comparison is one case. Evaluating performance requires many predictions and clearly defined measures, and the usefulness of a forecast may depend on the decision it supports.
The learning that lasts beyond weather
This method teaches a general rule of evidence: compare what was actually predicted with what was actually measured, under matching definitions. The same discipline helps students compare experimental readings, budgets and before–after results without moving the goalposts.
For related practice, see baseline readings versus control groups and measurement precision. Families can explore the eduKate Science learning guides and Education hub. The Clementi teaching reference shows the same diagnosis-to-independent-checking approach.