Real-World Mathematics: Weather, Climate, Averages, Anomalies and Forecast Uncertainty

Application of Mathematics in Real-World Usage · Guide 23 · BTT Mathematics Hub

A day can be cooler than normal while the month is warmer than normal. Two stations can have the same average rainfall but very different day-to-day variability. A 70% probability of rain is not a promise that 70% of the day will be wet. Weather and climate are full of averages, baselines, distributions and forecasts whose meaning depends on the denominator and time window.

This page is a Mathematics lesson, not a weather warning, travel instruction or climate-policy assessment. All temperatures, rainfall totals and forecast probabilities below are fictional teaching data unless a definition is linked to NOAA or WMO-related material. For real safety decisions, use current official meteorological information for the location and hazard involved.

The central distinction is between weather, which describes atmospheric conditions over relatively short times and places, and climate statistics, which describe distributions and typical conditions over longer periods. Mathematics connects individual observations to baselines, anomalies and probabilistic forecasts without turning uncertainty into certainty.

A climate normal is a reference baseline

NOAA’s National Centers for Environmental Information describes WMO climatological standard normals as products that summarise typical climate conditions and provide a benchmark for comparing current weather and forecasts. The current standard period widely used in the product set is 1991–2020.

Suppose a fictional location has a January normal mean temperature of 26.0°C. If January 2026 averages 27.1°C, the anomaly relative to that baseline is +1.1°C.

The anomaly does not mean every January day was 1.1°C warmer than normal. It compares one aggregated value with another aggregated reference.

Anomaly is observed value minus reference value

NOAA’s climate glossary defines anomaly as the difference between a value and its normal or long-term mean. Mathematically:

Anomaly = observation − reference.

If observed rainfall is 180 mm and normal rainfall is 220 mm, anomaly is −40 mm. As a percentage of the normal, the anomaly is −40/220×100≈−18.18%.

The absolute anomaly and percentage anomaly are different quantities. The percentage needs a nonzero reference denominator.

Changing the baseline changes the anomaly

A month has observed mean 27.0°C. Relative to a 26.0°C baseline, anomaly is +1.0°C. Relative to a 26.5°C baseline, it is +0.5°C.

The observation did not change. The reference did.

This is why anomaly charts must state their baseline period. Comparing anomalies built from different normals can create apparent differences that come partly from the reference choice.

Daily mean temperature is an average with a defined method

Suppose four equally weighted observations are 24, 28, 31 and 27°C. Their arithmetic mean is 27.5°C.

A different operational definition—such as averaging daily maximum and minimum—can produce a different number. The averaging method belongs to the quantity definition.

Therefore two “daily means” should not be compared until their construction is compatible.

Monthly means require appropriate weights when months have different lengths

Suppose January mean is 26°C and February mean is 28°C in a non-leap year. A simple mean gives 27°C, but a day-weighted mean across both months is (31×26+28×28)/59≈26.949°C.

The difference is small here, but the principle matters: averaging averages requires weights when the underlying sample counts differ.

The safest route is total weighted sum divided by total weight.

Rainfall totals add through time

Five fictional daily rainfall totals are 0, 12, 5, 30 and 3 mm. The five-day total is 50 mm.

The mean daily rainfall is 10 mm/day. Total and mean should not be swapped: 50 mm describes accumulation over five days, while 10 mm/day is a rate-like average over that interval.

If the period changes to ten days, a comparable total requires all ten observations; a comparable average still divides by the relevant number of days.

Rainfall intensity divides depth by duration

A fictional event produces 24 mm over 2 hours. Average event intensity is 12 mm/h.

If 20 mm fell in the first 30 minutes and 4 mm in the next 90 minutes, the event-average is still 12 mm/h, but the short-period intensities are very different.

An average over a long interval can hide brief peaks. Time resolution belongs to the measurement.

Spatial averages need area or representativeness weights

Suppose three station values are 100, 120 and 180 mm, but they represent 20%, 50% and 30% of a region respectively.

Weighted regional estimate is 0.2×100+0.5×120+0.3×180=134 mm.

The simple station mean is 133.33 mm. In this example the difference is small, but it can be much larger when station values and represented areas differ substantially.

A dry spell count is a rule-based classification

Define, purely for a classroom exercise, a dry day as rainfall below 1 mm. For values 0, 0.5, 2, 0, 0, 0.8, 5, the dry-day indicators are yes, yes, no, yes, yes, yes, no.

There are five dry days, with longest consecutive dry spell of three days.

If the threshold changes from 1 mm to 0.1 mm, the classification changes. Threshold-based climate indices depend on their exact rule.

A moving average smooths short-term variation

Take daily values 10, 14, 13, 17, 16. The three-day moving averages are (10+14+13)/3=12.33, then (14+13+17)/3=14.67, then (13+17+16)/3=15.33.

Smoothing reduces high-frequency variation but also shortens the output series and can delay or blur abrupt changes.

A moving average is a transformation of the original data, not an additional measurement.

Mean and variability answer different climate questions

Dataset A is 26,26,26,26,26. Dataset B is 22,24,26,28,30. Both means are 26°C.

A has zero range; B has range 8°C. The same mean can accompany very different variability.

When assessing consistency, extremes or threshold exceedance, spread matters alongside centre.

Standard deviation gives one scale of variation

For 22,24,26,28,30, mean is 26. Treating the five values as a complete finite population, squared deviations sum to 40 and population variance is 8. Standard deviation is √8≈2.828°C.

If these are a sample used to estimate a larger distribution, the n−1 convention would give a different sample standard deviation.

The convention and target population should be stated.

Percentiles describe position inside a distribution

Suppose sorted annual maxima are 30,31,32,33,34,35,36,37,38,39. The median is average of fifth and sixth values, 34.5.

Percentile definitions can vary in finite samples because interpolation rules differ. Therefore software settings can produce slightly different percentile values from the same dataset.

The conceptual meaning remains: a percentile locates a value relative to the distribution, not relative to physical zero.

A probability forecast is not a deterministic statement

Suppose a forecast assigns 70% probability to rain in a defined place and period. The event still has only two realised outcomes under that definition: rain or no rain.

The 70% describes uncertainty before the outcome is known. One realised dry day does not by itself prove the forecast was irrational, just as one rainy day does not prove a 30% forecast was perfect.

Probabilistic forecasts are evaluated across many cases by calibration and scoring, not by demanding that each percentage appear as a fraction of one day.

Probability categories must sum consistently

A three-category forecast might assign 50% above normal, 30% near normal and 20% below normal. The probabilities sum to 100%.

If the forecast changes above-normal probability to 60% while the other categories remain 30% and 20%, the result sums to 110% and is invalid as a complete mutually exclusive probability distribution.

NOAA Climate Prediction Center outlooks use probability categories relative to climatology; the exact operational method belongs to the forecast product.

Expected value summarises a probability distribution

Suppose a teaching forecast for rainfall amount uses 0 mm with probability 0.2, 10 mm with 0.5 and 30 mm with 0.3. Expected rainfall is 0×0.2+10×0.5+30×0.3=14 mm.

The expected value need not be one of the actual possible outcomes. In this model the realised rainfall is 0,10 or30 mm, never 14.

Expected value is a probability-weighted average over repeated comparable situations.

Brier score evaluates binary probability forecasts

For a binary event, Brier score for one case is (p−o)², where p is forecast probability and o is 1 if the event occurs, 0 otherwise.

A 70% rain forecast followed by rain gives (0.7−1)²=0.09. If no rain occurs, score is (0.7−0)²=0.49. Lower is better.

Across many forecasts, average Brier score evaluates probabilistic accuracy. It does not alone reveal whether errors come from bias, poor calibration or lack of resolution.

Ensemble spread is not automatically a confidence interval

Suppose five model members predict 20,22,24,25 and29 mm. Ensemble mean is 24 mm and range is 9 mm.

The member range shows spread among those model realisations. It is not automatically a 100% prediction interval for the atmosphere.

Interpretation depends on how the ensemble was generated and calibrated. The Mathematics should label member spread as member spread unless a validated probabilistic interpretation is supplied.

Forecast bias is average forecast error with a sign convention

Define error as forecast minus observed. Forecasts 10,12,15 and observations 9,13,14 give errors +1,−1,+1.

Mean error is 1/3≈+0.333. Mean absolute error is (1+1+1)/3=1.

A small mean error can hide larger cancelling errors, so bias and typical error magnitude should be reported separately.

Temperature trends depend on the regression model

Suppose annual anomaly values at years x=0,1,2,3 are 0.1,0.2,0.4,0.5°C. A least-squares line has slope 0.14°C per year and intercept 0.09°C for this tiny fictional dataset.

The slope summarises the linear component. It does not prove the process is truly linear or that the same slope will continue into the future.

Trend inference in real climate data must consider length, autocorrelation, measurement quality and model assumptions.

Autocorrelation reduces the information in closely related observations

Daily temperatures on consecutive days are often more similar than independent random draws. Treating every day as statistically independent can overstate effective sample size.

NOAA’s climate glossary defines autocorrelation as correlation between time-series values and values at a fixed lag.

The mathematical consequence is important: formulas derived for independent data need adjustment when observations are serially correlated.

A threshold exceedance count depends on the threshold

For daily maxima 30,31,33,35,29,34, define a hot day in this teaching exercise as ≥33°C. Four days qualify: 33,35,34 and one more? Check carefully: the qualifying values are 33,35,34, so there are three.

This deliberate check illustrates why threshold classifications should be enumerated rather than inferred from a quick glance.

If the threshold is changed to ≥34°C, only 35 and34 qualify, giving two.

Return periods are long-run frequency summaries, not schedules

In a simplified independent annual model, an event with annual exceedance probability p=0.01 has return period 1/p=100 years.

That does not mean it occurs exactly once every hundred years. The probability of at least one occurrence in 100 independent years is 1−0.99¹⁰⁰≈63.4%.

Real extreme-value analysis is more sophisticated; this example only demonstrates how probability and return period relate under a simplified stationary model.

Forecast uncertainty should change the wording of conclusions

Suppose a modelled temperature range is 31–34°C. A threshold of 33°C lies inside the range.

The model supports “33°C is possible under the supplied range,” but not “temperature will exceed 33°C” or “will remain below 33°C.”

Mathematical uncertainty is not merely an error bar to ignore; it changes which statements are logically justified.

A complete climate-statistics report states time, place and baseline

For an anomaly, state the observed period and normal period. For an average, state weighting and missing-data treatment. For rainfall, state accumulation interval. For probability, define the event and verification period.

Then label outputs as observation, climatology, deterministic model, ensemble statistic or calibrated probability forecast.

The Mathematics becomes clearer when these objects are prevented from collapsing into one generic word such as “prediction.”

Practice: twenty weather and climate questions

1. Observed monthly mean is 27.1°C and normal is 26.0°C. Find anomaly.
2. Observed rainfall is 180 mm and normal is 220 mm. Find absolute and percentage anomaly.
3. Find mean of 24,28,31,27°C.
4. Find day-weighted mean of January 26°C for 31 days and February 28°C for 28 days.
5. Find five-day rainfall total for 0,12,5,30,3 mm.
6. Find mean daily rainfall for question 5.
7. Find average intensity for 24 mm over 2 hours.
8. Find weighted regional rainfall for 100,120,180 mm with weights 0.2,0.5,0.3.
9. For 0,0.5,2,0,0,0.8,5 mm and dry threshold <1 mm, find dry-day count and longest dry spell.
10. Find three-day moving averages of 10,14,13,17,16.
11. Find population SD of 22,24,26,28,30.
12. Forecast categories are 50%,30%,20%. Check sum.
13. Find expected rainfall for outcomes 0,10,30 mm with probabilities 0.2,0.5,0.3.
14. Find Brier score for 70% forecast when event occurs.
15. Find Brier score for same forecast when event does not occur.
16. Find ensemble mean and range for 20,22,24,25,29 mm.
17. Forecasts 10,12,15 and observations 9,13,14. Find mean error and MAE using forecast−observed.
18. For annual exceedance probability 0.01, find nominal return period.
19. Find probability of at least one occurrence in 100 independent years.
20. Why can a positive monthly temperature anomaly coexist with cooler-than-normal individual days?

Worked answers

1. +1.1°C.
2. −40 mm and about −18.18%.
3. 27.5°C.
4. About 26.949°C.
5. 50 mm.
6. 10 mm/day.
7. 12 mm/h.
8. 134 mm.
9. Five dry days; longest spell three days.
10. 12.33, 14.67 and 15.33.
11. √8≈2.828°C.
12. 100%.
13. 14 mm.
14. 0.09.
15. 0.49.
16. Mean 24 mm; range 9 mm.
17. Mean error +0.333; MAE 1.
18. 100 years.
19. About 63.4%.
20. Because the monthly mean aggregates many days. Some negative daily anomalies can coexist with a positive average if other days are sufficiently above normal.

Sources and connected applications

For climatological normals, see NOAA NCEI: WMO Climate Normals. For anomaly and autocorrelation definitions, see the NOAA Physical Sciences Laboratory Climate Glossary. For probabilistic outlook context, see the NOAA Climate Prediction Center glossary. The numerical datasets and forecast cases here are original teaching examples.

Continue with Environmental Monitoring, Emissions, Concentrations and Mass Balance; Population, Demography, Growth and Dependency Ratios; and Astronomy, Scale, Orbits, Angles and Observation. Return to the BTT Mathematics Hub.