Historical Weather Accuracy: What You Can Trust

Separate measurement error from representativeness

Measurement error concerns the instrument and recording process: calibration offsets, damaged shielding, clock errors, or transcription mistakes. Representativeness concerns whether a valid reading describes your target. A correctly functioning thermometer on an open rural field need not match a nearby dense urban street at night.

Processing introduces another layer. A daily average calculated from hourly samples differs from an average of daily maximum and minimum. An accumulation assigned to a morning observation date may not match a midnight-based calendar day. Some apparent disagreements therefore reflect definitions rather than defective measurements.

Precision is not accuracy. Showing 18.37 °C can imply more certainty than the source supports. Rounding a result sensibly does not solve exposure problems, but it avoids presenting extra decimal places as evidence of better knowledge.

Understand what each source represents

NOAA GHCN-Daily integrates station summaries subjected to common quality-assurance reviews. Its documentation and inventories help identify variables, coverage, and flags. Quality control screens suspect observations; it does not make an airport's precipitation gauge a measurement of every neighbourhood.

ECMWF ERA5 is reanalysis: observations and a numerical model jointly estimate atmospheric conditions. Its spatial continuity is useful where stations are sparse, but a grid estimate can smooth local showers, terrain effects, and extremes. The appropriate comparison is not “real versus fake”; it is a point observation versus a model-based spatial representation.

The Open-Meteo archive documentation identifies its datasets, elevation handling, and variable definitions. A coordinate request does not establish a sensor at that coordinate. Record the selected model and returned location information when evaluating its output.

Align values before testing agreement

1. Use the same units and distinguish temperature from apparent temperature.
2. Match the timezone and the exact interval being summarized.
3. Compare like statistics: mean with mean, maximum with maximum.
4. Remove or separately review provider-flagged values using a declared rule.
5. Use only paired dates when calculating comparison metrics.
6. Inspect station elevation and model terrain before interpreting offsets.

Do not fill missing observations with model values and then test the model against that blended series. That makes the evaluation partly compare the product with itself. Keep validation observations separate, and disclose whether the reanalysis may already have assimilated information from the same observing network.

Illustrative example: bias can hide larger errors

Illustrative comparison only; these temperatures are invented, not real observed data from a station or model. Consider four aligned daily mean temperatures. Define each difference as estimate minus station reference. Treating the station as a reference here does not imply that it is error-free.

Mean bias is (1 − 1 + 2 + 0) ÷ 4 = +0.5 °C. Mean absolute error is (1 + 1 + 2 + 0) ÷ 4 = 1.0 °C. Root mean square error is the square root of (1 + 1 + 4 + 0) ÷ 4, approximately 1.22 °C. The last measure gives larger errors extra weight.

A modest average bias does not mean every day is close: positive and negative differences cancel. Subtracting 0.5 °C from every estimate would correct this sample's mean bias, not prove accurate daily reconstruction. Four pairs are far too few to establish performance across seasons or extremes.

Check the conditions that average metrics conceal

Evaluate summer and winter separately if the intended use spans both. Inspect warm nights, heavy-rain days, and windy conditions where relevant. A model can perform well on ordinary temperatures while missing a short heat peak. A rain series with many dry days can look close on average while poorly representing the few storms that matter.

For rain occurrence, use a declared threshold and count hits, misses, and false alarms. For amounts, examine accumulation windows and whether the reference gauge is representative. Avoid a relative error calculation on a zero reference amount; division by zero is undefined. Percentage errors are also unstable near tiny totals.

Comparing two websites is not necessarily independent corroboration. They may redistribute the same station archive or the same reanalysis. Trace the underlying dataset before claiming that agreement between services independently confirms an event.

Set validation rules before inspecting the results. Choosing only nearby stations that agree, deleting difficult storm days, or tuning a correction and evaluating it on the same dates can make performance look artificially strong. Reserve a separate period for checking any fitted adjustment, and retain an audit trail of exclusions. Look for gaps concentrated during severe conditions: complete pairs may disproportionately represent easy weather. Report both the retained sample and rejected observations, because a tidy error statistic cannot reveal selective sampling by itself.

Write confidence as a specific limitation

Prefer “regional temperature context from a gridded reconstruction” to “exact weather at this address.” Prefer “daily precipitation at station X” to “the whole city received this amount.” Include the usable sample count, missing-data policy, and any unresolved time or exposure mismatch. There is no universal error margin valid for all historical variables and places.

Use the comparison tool to explore differences, the date tool to locate an event, and methodology to establish source type. For insurance, engineering, or disputed records, seek the relevant official archive and qualified interpretation rather than treating an educational summary as certified evidence.

Related articles

• Station Distance and Data Quality
• Common Weather Data Mistakes
• How Historical Weather Data Is Collected

References

• NOAA: GHCN-Daily quality assurance and metadata
• ECMWF: ERA5 product description
• Open-Meteo: historical datasets, units, and time definitions

About the author

WeatherRecall editorial project compiles educational explanations from listed public documentation. It is not a credentialed meteorologist and has not conducted original field research or a validation campaign for this article. The error calculations are hypothetical demonstrations, not measured product performance.