Must a time series be numerical?

Many common methods use numerical values, but time-indexed observations can also describe categories or events.

Must observations be equally spaced?

No. Irregularly observed series exist, although analysis methods must respect their timing.

Is a time series a forecast?

No. It is data indexed by time; a forecast is a prediction of future values or events.

AI Glossary / Time series and signal processing / Advanced

Time series

Definition

A time series is a sequence of observations associated with times or time intervals. Its ordering is part of the data: measurements taken earlier and later can have different relationships and meanings. Examples include hourly sensor readings, daily order counts and monthly demand. A time series may contain one variable or several aligned variables, and its observation times need not always be equally spaced.

What it means

A list of values becomes useful for temporal analysis when each value has a clear measurement definition and time reference. Twenty orders during an hour is different from twenty orders waiting at the end of that hour. The first is activity accumulated over an interval; the second is a snapshot. Combining or comparing observations requires preserving that distinction, along with units, coverage and collection practices.

How it works

Record a time index, the measured values and metadata explaining what each observation represents. Check ordering, duplicate timestamps, gaps and changes in definitions. Determine whether observations are totals, rates, averages or point-in-time measurements before aggregation.
Analysis can then examine changes, recurring patterns and dependence between observations. Forecasting is one possible use, but a time series can also support historical description, monitoring or anomaly investigation. Temporal relationships are not guaranteed merely because data have timestamps; they must be assessed in the context of the process and collection method.

Three practical examples

A retailer records orders completed each day. It distinguishes days with no orders from days when the reporting system failed, since a zero and a missing value imply different evidence.
A building records temperature every ten minutes. If a sensor stops reporting for an hour, the next reading is not simply another ten-minute step unless the gap is explicitly handled.
A service team records open cases at the end of each week. Adding those snapshots does not count unique cases handled, because the same unresolved case can appear in several weeks.

Think of it this way

A time series is like a film rather than a shuffled box of photographs. The individual frames matter, but their order and spacing explain how the scene developed. Sorting the frames by brightness might preserve their collection of colors while destroying the story of what happened first and what followed.

Worked example

Construct six monthly order totals for Series A: 10, 20, 30, 40, 50 and 60. Construct Series B from the same values in a different order: 10, 50, 20, 40, 30 and 60. Both have total 210, mean 35 and median 35. Their distributions of individual values are identical, but their time patterns differ.
A's five month-to-month changes are 10, 10, 10, 10 and 10. B's changes are 40, -30, 20, -10 and 30. Both rise by 50 from the first month to the last, yet total absolute month-to-month change is 50 for A and 130 for B. A summary that ignores ordering misses that difference.
If the first three months form one quarter and the next three another, A's quarterly totals are 60 and 150. B's are 80 and 130. Aggregation changes the temporal view even though both six-month totals remain 210. These sums are meaningful because the stipulated observations are nonoverlapping order totals. The same addition would not represent end-of-quarter inventory if the numbers were stock snapshots.
Now consider a separate six-day series: 12, 14, missing, 18, 16 and 20. The five known values sum to 80. Their observed-day mean is 80/5 = 16. Dividing by six gives 13.3333, implicitly treating the missing observation as zero. Neither calculation establishes the true six-day mean. If the missing value were 10, that mean would be 15; if it were 22, the mean would be 17.
Finally, suppose two reporting intervals contain 30 orders over two hours and 20 orders over one hour. Their individual rates are 15 and 20 orders per hour. The combined rate is 50/3 = 16.6667 orders per hour, not their unweighted average of 17.5. The interval lengths determine the appropriate weighting.

Illustrative case study

Fictional Cedar Services builds a reporting dataset from daily case records. Its first export drops dates without a report, making later rows appear adjacent in time. A reviewer adds an explicit daily index and separates confirmed zero activity from unavailable records.
The team also labels completed-case totals separately from open-case snapshots. It checks any weekly summary against the meaning of its underlying measurements before feeding it into a forecasting exercise. This scenario illustrates preparation choices, not proof that the resulting data are complete or that an AI model will predict them accurately.

Common misconceptions and mistakes

A timestamp does not automatically establish regular spacing, comparable coverage or causal order. Event time and reporting time can differ. Sorting observations by value destroys their temporal sequence, while replacing gaps with zero changes their meaning. A trend in a short record is not evidence that it will continue. Combining series from different entities also requires identifiers; otherwise observations from separate processes can be mistaken for one continuous history.

Business use

Define the measurement, time zone, interval boundaries and aggregation rule before choosing an analysis tool. Record revisions and distinguish information available at a decision time from values corrected later. For forecasting evaluation, recreate the intended information cutoff so future observations do not enter earlier predictions. Clear definitions often matter as much as model choice because a sophisticated model cannot repair an undocumented change in what the target measures.

Check your understanding

Why do Series A and B require different temporal descriptions despite having the same mean and median?

Answer: Their values occur in different orders. A rises steadily in the construction, while B alternates increases and decreases. Their month-to-month changes and quarterly totals therefore differ.

Related terms

Flow series
Stock series
Missing data
Autocorrelation
Forecasting
Time index
sampling interval
temporal aggregation

Learn next

For Series B, replace month six's value of 60 with 50. Its total becomes 200, mean becomes 33.3333, final month-to-month change becomes 20 and total absolute change becomes 120. Identify which summaries depend on time order and which depend only on the collection of values.

Sources

Australian Bureau of Statistics, Statistical terms and concepts glossary, Time series, Flow series and Stock series: https://www.abs.gov.au/statistics/understanding-statistics/statistical-terms-and-concepts/statistical-terms-and-concepts-glossary Glossary of Forecasting Terms: https://robjhyndman.com/mwh3/FG4.pdf All numerical series, reporting intervals and the Cedar Services scenario are original teaching constructions. Missing values are left explicitly unknown rather than silently imputed.

Previous: Time origin Next: Time series analysis