AI Glossary / Data science and analytics / Beginner
Decile
Definition
A decile is a quantile associated with dividing an ordered distribution into ten parts. The nine internal decile cut points correspond to the 10th, 20th, through 90th percentiles. In everyday reporting, decile can also mean one of the ten resulting groups, such as the top decile. Reports should distinguish the boundary value from the group of observations.
What it means
Decile groups describe relative position in a defined distribution. The top decile contains the highest-ranked tenth under a stated grouping convention, not necessarily values above 90 on an absolute scale. Equal-sized groups need not span equal numerical ranges. In finite datasets, ties, weights, and calculation conventions can complicate an exact division into ten equally sized groups, so the method needs to be documented.
How it works
Choose the population, variable, weighting, and ordering. Calculate cut points or assign observations to rank-based groups using an explicit convention. State how ties and missing values are handled and whether ranks run from low to high or high to low. When comparing periods or datasets, decide whether boundaries are recalculated each time or held fixed. Recalculated deciles describe relative position within each distribution; fixed boundaries instead track values against an earlier reference. These approaches answer different questions.
Three practical examples
- Income reporting: statistical summaries compare positions in a defined income distribution, with the population and income measure clearly specified.
- Model evaluation: analysts rank cases by predicted score and inspect observed outcomes in groups representing tenths of the ranked sample.
- Operations: a team studies the slowest tenth of completed requests while separately examining the actual duration threshold defining that group.
Think of it this way
Think of lining up 100 items by a measurement and dividing the line into ten groups of ten. Each group contains the same number of items, but the numerical gaps between neighboring items may be very different. The analogy explains rank grouping. Real datasets may have ties, survey weights, or sizes that do not divide evenly into ten.
Worked example
Suppose a fictional evaluation contains 100 cases with distinct prediction scores. Rank them from highest to lowest and put the first 10 in the top score decile. Six of those 10 cases have the outcome of interest, while 20 of all 100 cases have that outcome. The top-decile observed rate is 6/10 = 60%, compared with an overall rate of 20%.
The top decile contains 10% of cases but captures 6/20 = 30% of the observed outcomes. Those are different percentages with different denominators. The result does not prove the model's probabilities are calibrated, and it does not mean the top group's scores exceed 0.9. Its score boundary depends on the sample and ranking. Independent evaluation is still needed before using the ranking operationally.
Illustrative case study
In a fictional service organization, managers celebrate that the slowest decile of requests is smaller after a process change. Analysts clarify that a decile defined by ranking will still contain approximately a tenth of eligible requests. What may change is the duration boundary or the durations within that group. They report the 90th-percentile duration alongside the definition of eligible requests and the full time distribution. The scenario illustrates why a rank label alone cannot show absolute improvement and why a changing denominator must remain visible.
Common misconceptions and mistakes
Deciles are not equal-width bins. The top decile is not the top ten observations unless the relevant sample contains 100 equally weighted observations under the chosen grouping rule. A decile cut point is not the same as the average value within that group. Software can return different sample quantiles because interpolation conventions differ. Ties may prevent both equal group sizes and identical treatment of equal values at the same time; the analyst must choose and explain a policy rather than hide the conflict.
Business use
Businesses use deciles to make ranked distributions manageable, but should retain the underlying scale and population context. In AI evaluation, decile charts can show ranking behavior, concentration, or calibration patterns depending on what is plotted. Name the metric and denominator precisely. When deploying a score threshold, decide whether a fixed fraction or a fixed score is intended, because workload and risk implications differ. Decile summaries can conceal substantial variation within groups, so they should not replace more detailed checks when those differences matter.
People also ask
Are there nine or ten deciles?
There are nine internal cut points dividing a distribution into ten groups; everyday language may call those groups deciles.
Must each group cover the same numerical range?
No. Equal rank shares can correspond to very unequal value ranges.
Will different software always return identical cut points?
No. Sample quantile definitions and tie handling can differ, so record the method used.
Check your understanding
A model's top decile has scores between 0.32 and 0.58. Is that impossible because a top decile must be above 0.9?
Answer: No. Top decile means highest-ranked tenth of the defined sample, not an absolute probability range. Scores and rank positions are distinct.
Related terms
Learn next
rank a small dataset and compare equal-count groups with equal-width intervals. Introduce tied values at a boundary and decide how to handle them transparently. Then compare a boundary recalculated each period with a fixed reference threshold to see how their interpretations differ.
More in Data science and analytics
- Data Transformation - Data transformation is the conversion of data from one representation, structure, or set of values into another according to defined rules.
- Data visualization - Data visualization represents data through graphical elements such as position, length, color, shape, or size.
- Data wrangling - Data wrangling is the work of turning available data into a form suitable for a particular analysis or application.
- Deep Copy - A deep copy creates a new compound object and recursively copies its contained objects according to the language or library's copying rules.
- Deep learning - Deep learning is a branch of machine learning that learns representations through multiple layers of computation.
- Degree-Rips bifiltration - A degree-Rips bifiltration is a two-parameter construction in topological data analysis that examines a point cloud using both distance scale and local...
Sources
Sources: Statistics Canada's decile-group definition illustrates rank-based population grouping: https://www12.statcan.gc.ca/census-recensement/2016/ref/dict/pop166-eng.cfm . Its census dictionary provides additional decile context: https://www12.statcan.gc.ca/census-recensement/2021/ref/dict/98-301-x2021001-eng.pdf . Relevant primary definition excerpts were consulted; no income values or population findings are adopted. All prediction-score and service examples are original illustrations, and implementation-specific quantile conventions require explicit documentation.
Previous: Decentralized planning Next: Decision analysis