What makes data numerical?

Values on a number line where arithmetic is meaningful: magnitudes, counts and measurements, unlike labels or codes.

What is the difference between discrete and continuous?

Discrete numbers count indivisible units like orders; continuous numbers measure, like weight, and can be infinitely fine.

Why scale numeric features?

So models that compare magnitudes judge relevance rather than units, since raw scales make large-unit fields dominate.

Are numeric IDs numerical data?

No: they are labels stored as digits, and treating them as quantities lets models learn meaningless orderings.

Not directly: dates become numerical through engineering, as ages, durations or day counts, and the conversion choice is a modeling decision.

AI Glossary / AI and machine learning / Beginner

Numerical data

Definition

Numerical data is data that lives on a number line: quantities with magnitude and order, like revenue, temperature, age or count, where arithmetic means what it says.

What it means

The defining test is whether arithmetic makes sense: the average of two revenues means something, the average of two phone numbers does not, though both are stored as digits.

Numerical data splits into continuous values, which can be infinitely fine, like weight, and discrete counts, like number of orders, and both answer the same families of questions: how much, how many, how does it trend.

Numbers still carry choices: units, time zones, currencies and rounding are conventions, and a model reads the digits exactly as stored, with no idea what they mean.

Most models want numbers and only numbers, which is why categorical data gets encoded into numeric form before training, and why numeric features get scaled so their magnitudes compare fairly.

For a business, numerical data is the easy half of modeling: it feeds models directly, but it is also where quiet errors live, in units, outliers and the difference between a code and a quantity.

How it works

Verify each numeric field's unit and currency, and pin conversions before any comparison or modeling.

Check distributions for outliers and impossible values, since a single mis-keyed zero can bend every average.

Scale features before distance-based or regularized models, so magnitude reflects relevance rather than units.

Separate true quantities from numeric codes: IDs, postcodes and phone numbers are digits, not numbers.

Track drift in distributions over time, because a shifting numeric field quietly changes what the model sees.

Decide the handling of skewed fields early: incomes and counts often span orders of magnitude, and a log transform can matter more than the model choice.

Three practical examples

A revenue field mixing three currencies until a conversion step reveals the apparent growth was an exchange-rate shadow.
An age field containing 999 as a missing-data code, inflating averages until the convention is discovered and handled.
A delivery-time model improving after features are standardized, so days and kilometers finally compare on fair scales.

Think of it this way

Think of a market scale: it reads whatever sits on it, in the units it was set to. Trust the reading only after checking the dial, the tare and whether the scale was even on.

Worked example

the 999 customers

Suppose an analyst finds the customer base's average age is 41.2, up from 40.8 last year. The figures are illustrative.

The rise traces to a new form storing 999 for skipped ages: three per cent of records are centuries old, and the 'aging customer base' is a data convention, not a trend.

Recomputed with the codes removed, the average is 40.9, and the trend story dies, along with the campaign brief built on it.

The field was numerical, the storage was clean, and the meaning was broken: numeric data rewards arithmetic only after its conventions are audited.

Illustrative case study

A fictional lender fed an ID column into its risk model as a numeric feature, and the model dutifully learned that higher IDs were riskier, because IDs grew with time and defaults had risen with a rate change.

The model passed validation on shuffled data, failed forward in time, and the post-mortem found a serial number standing in for the economy.

The invented lesson: digits are not quantities. Audit every numeric field for what its numbers mean before letting arithmetic at them.

The repair was a type audit: every field classified as quantity or code, the codes removed from numeric features, and the model retrained on what the digits meant.

Common misconceptions and mistakes

Digits mean numbers. IDs, codes and postcodes are labels made of digits, and arithmetic on them is nonsense.
Numbers need no preparation. Units, outliers, scaling and codes all decide what the arithmetic actually says.
Numerical data is objective. Every number arrives through a convention, a unit, an instrument and a process, each of which can lie.

Business use

Audit units and codes on every numeric field before analysis, especially after system changes.

Exclude identifiers and codes from numeric features, and say why in the data dictionary.

Scale numeric features for models that compare magnitudes, and record the scaling for reproduction.

Watch distributions over time, because a drifting field changes the model's world without announcing itself.

Check your understanding

Why does the digit-versus-number distinction matter for models?

Answer: Because models apply arithmetic to whatever they are fed, and arithmetic is only meaningful on true quantities: codes and IDs stored as digits get read as magnitudes, letting the model invent orderings and effects that exist nowhere but in the storage format.

Related terms

Learn next

Be Productive Using AI builds the data habits that keep arithmetic honest: https://aiforbusinessandlife.com/academy/be-productive-using-ai. [C1]

Sources

[S12] Aalto University, Dictionary of Machine Learning; [S14] KNIME, Data Science Glossary. Full URLs are in the source ledger. Verified 11 October 2026.

Previous: Numeric variable Next: Nyquist frequency