How ScoutDNA calculates what it shows

ScoutDNA consumes football data through a provider interface. No part of the product reads a provider format directly, which means a licensed feed can replace the current source without changing how anything is calculated or displayed.

Where no licensed source is configured, the product serves clearly labelled sample data. Sample players, clubs and competitions are invented. They are not anonymised real records, and no figure in the sample dataset describes a real person or organisation.

The product never fabricates a statistic. When a source does not supply a metric, the metric is absent, the affected score is calculated from what remains, and the coverage is reported alongside the result.

Raw season totals are not comparable. A midfielder with 3000 minutes will out-total one with 900 regardless of quality, and a pass completion rate cannot be divided by minutes at all.

• Counting metrics are converted to a per 90 minute rate.
• Rates and percentages are used as reported, never divided by minutes.
• A player with no minutes has no rate. They are excluded rather than recorded as zero, because zero would rank them alongside a player who played a full season and produced nothing.

A percentile only means something relative to a group. ScoutDNA ranks each metric within the combination of position group, competition and season.

Position alone would compare a second division midfielder with a top flight one. Competition alone would compare a goalkeeper with a winger. A second percentile is stored against the position group across all competitions, for cross league context.

• Ties take the midpoint, so a group of identical values sits at the 50th percentile rather than at 0 or 100.
• Metrics where fewer is better, such as fouls or dispossessions, are inverted, so a low count reads as a high percentile.
• A peer group smaller than five players is not ranked at all. A percentile against four peers carries no information.

Percentiles are recalculated by a scheduled run, not inside a page request, and every stored value records the size of the group it was measured against.

A role aware weighted average of a player's percentiles, expressed on the same 0 to 100 scale. It summarises measured output within a peer group. It is not a rating of ability, and it does not account for tactical instruction, team quality or anything a camera would show.

Components are missing when the source supplied no metric for them. The remaining weights are renormalised, so a player is scored on what is known rather than penalised with a zero for data that was never collected. The coverage is reported with the score.

A DNA profile describes a role as a set of weighted metrics, each with the percentile that role tends to produce. Matching measures how close a player sits to those expectations within their own peer group.

The measure is distance from a target, not "more is better". A deep-lying playmaker profile expects high passing volume; a player at the 99th percentile for tackles is a different kind of midfielder, not a better playmaker. Exceeding a target costs far less than falling short of it, but it is not free.

• A metric with no data is dropped and the remaining weights renormalised. It is never scored as zero.
• A profile can mark a metric as required. A player with no data for it is excluded rather than scored on a partial profile.
• Each profile declares the positions it applies to. A centre-back is not scored against a full-back profile.
• Every profile sets a minimum minutes threshold. Below it, the player is not scored.

Statistical similarity is not a prediction of career outcome, and it does not establish that two players occupy the same tactical role. A high match means measured output resembles a defined pattern. What that is worth is a scouting judgement, not a model output.

These are called signals rather than scores because each one measures something observable and says nothing about what a player will become.

• Upside combines remaining development runway by age, output relative to age, season on season trend and playing time trajectory. A trend is measured against the player's most recent earlier season, and the explanation names which season that was.
• Hidden Value compares output against market attention. Without a licensed market value source the valuation component is absent, and the signal reports that rather than estimating a value.
• Breakout measures acceleration rather than level. A settled high performer scores low, which is correct: they have already broken out.

Confidence is a label, not a number. A figure such as "97.4% confident" implies a statistical guarantee this model cannot make.

Confidence is the weakest of its inputs, not their average. A player with 3000 minutes but 40% metric coverage is a limited confidence case, because the thin part is the part that limits what can be said.

30 metrics are defined. A source that does not supply one leaves it absent rather than zero. Metrics marked below as inverted are ones where a lower count is the better outcome, and their percentiles are flipped accordingly.

Every stored score records the model version that produced it. Changing a weight does not silently rewrite history: results calculated under an earlier version keep their version tag, so a shortlist built months ago can still be explained.

Current scoring model 1.0.0. Current DNA model 1.0.0.

Stating the limits is part of the method, not a disclaimer appended to it.

• Tactical instruction. A midfielder told to sit deep will produce different numbers in the same body.
• Team quality and possession share, beyond the competition context adjustment.
• Anything visible only on video: body shape, scanning, decision making under pressure, temperament.
• Injury history, contract position and availability, unless a licensed source supplies them.
• Fit with a specific squad, manager or dressing room.

ScoutDNA narrows a market to a shortlist worth watching. It does not replace watching them.