A radar chart can make the same player look better or worse depending on the scale. Use percentiles for rank, z-scores for distance from average, and raw numbers only when the scale is fixed and explained.

Scouts, analysts, and coaches use radar charts because they make a player profile easy to scan. The risk is that the shape can change even when the player’s output has not changed. A winger can look elite, average, or balanced depending on who he is compared with and where each axis starts and ends.

This piece is about the hidden choices behind those charts. It looks at peer groups, percentiles, z-scores, raw numbers, min-max scaling, fixed boundaries, and disclosure. Those choices matter because a recruitment meeting can treat the shape as evidence about the player, when part of the shape may really come from the chart design.

Start with who the player is being compared against

A radar number only means something inside a comparison group. A full-back ranked against top-five-league full-backs is being judged against a different standard from a full-back ranked against every player in his league, every wide defender in a database, or players in his age band.

Opta’s player radars use 15 years of top-five European league seasons, more than 24,000 outfield player performances, and a 1,350-minute threshold. In that dataset, Bukayo Saka’s 92 rating for chances created meant his 2.1 chances created per 90 ranked above 92% of forwards in the benchmark group.

Soccermatics’ statistical scouting guide says percentile rankings should compare players in a similar position in the same league. Its radar example ranks Mohamed Salah against comparable players and warns that radar areas can distort the data because the filled area is not proportional to the percentile.

Hudl StatsBomb’s 2023 radar update says its boundary values represent the top and bottom 5% of production by position group. The same update says metrics are calculated per 90 minutes, with defensive metrics possession-adjusted. Before anyone reads the shape, those choices have already decided what the shape can mean.

Use percentiles when the question is rank

Percentiles are the easiest scale for most public player profiles. They answer a simple question: how does this player rank against similar players? A 0-100 scale also lets different metrics sit on the same chart without asking the reader to know every original unit.

The trade-off is that percentiles hide distance. The gap between the 80th and 90th percentile can be small in raw output on one metric and large on another. If the reader needs both rank and production, show the raw per-90 number beside the percentile. Cannon Stats’ radar explainer uses that structure by pairing the radar with raw values below the graphic.

Percentiles are useful when the chart needs to be read quickly. They are weaker when the decision depends on how far above average the player actually is.

Use z-scores when the question is distance from average

A z-score states how far a value is from the mean in standard deviation units. In football terms, it can show whether a player is slightly above the group or genuinely unusual.

scikit-learn’s StandardScaler documentation describes the standard process: compute the mean and standard deviation on a training set, then apply the same transformation to later data. For a player radar, the same rule applies: fit the scale on the reference group and keep it fixed.

Z-scores are usually better for analysts than for quick public graphics. Negative values, standard deviations, and distribution shape all need explaining. A percentile is usually easier for a recruitment slide or public article.

Be careful with raw numbers and min-max scales

Raw numbers can be misleading when one radar contains different football units. Expected assists, duel win percentage, touches in the box, and fouls are not measured the same way. If they are forced onto one shared visual scale, the metric with the larger numbers can dominate the shape.

Min-max scaling rescales each value from the observed minimum and maximum for that metric. That can also mislead. One extreme player can stretch the scale and make everyone else look closer together. If the minimum and maximum change between charts, the same output can land in a different place.

Raw or min-max radars need fixed endpoints. Trimmed endpoints, such as 5th-to-95th percentile boundaries, can keep the chart readable while reducing the effect of one outlier season. The endpoints should be disclosed because they affect the evidence.

Use the same scale when comparing two players

Two player radars can only be compared cleanly when the axes use the same limits. StatsBomb’s radar explainer describes boundaries based on the top and bottom 5% of production by position across top-five European league data. mplsoccer’s Radar class makes the same idea explicit by requiring low and high values for every metric, plus a lower_is_better option for reversed stats.

Small multiples follow the same rule. BetterEvaluation’s guidance says mini-graphs need the same scale for viewers to compare them. If two full-backs are on one recruitment slide, their charts should not quietly use different boundaries.

Keep the chart small enough to read

Opta says its radar templates are usually limited to nine slices, grouped into three themes, with correlations checked so the chart does not repeat the same information through several metrics. Highcharts’ radar guidance recommends 5-8 axes and no more than 4-5 polygons.

One clean player profile is more useful than a crowded shape. If the chart needs 15 metrics, three overlays, and long labels, split it into smaller views or put the detail in a table.

Tell the reader what changed the shape

Lower-is-better metrics have to be handled consistently. Cannon Stats says turnovers, dribbled past, and fouls committed are reversed so that farther from the centre still reads as better. If that reversal is hidden, the reader cannot tell what the shape means.

Radar metrics can also describe role and team context as much as quality. Hudl StatsBomb’s update warns that centre-back outputs are tied to team tactical style and should be read with caution.

Duan, Tong, Sutton, Asch, Chu, Schmid, and Chen’s 2023 paper notes that radar chart connected regions can be misused because area depends on the ordering of axes. The filled shape should not be treated as a total player score unless the chart was built and explained for that purpose.

A published radar should name the sample, season, league level, position group, minutes threshold, per-90 or possession adjustment, scale type, fixed boundaries, lower-is-better reversals, and raw values when rank alone is not enough. If those choices are missing, the chart may still look clear, but the reader cannot check what it is really showing.