How we decide that a number is outside the norm

A full-back runs twice as far as a centre-back, and a goalkeeper has nothing to do with either. Comparing a player against the squad average shows the difference between positions dressed up as a difference in load: the number is true, the question it answers is another one.

That is why the default reference is the personal norm — that athlete’s own history, not the team’s — and the order is fixed: themselves first, then their position group if it is large enough, never the team average as the yardstick for a judgement (traffic light, alerts, readiness). When it falls back on the group, or in profiles on the team because the group is too small, the program says so.

In football the week turns around the match, and every day of the microcycle carries a different expected load by design: MD-3 is the hard day, MD-1 is the pre-match session, MD+1 is recovery. The calendar day, on its own, says nothing — a Tuesday can be MD-4 or MD+2, and those are two different worlds.

We measured it on the simulated season: within the same MD, load variability drops by 37% on distance and by 47% on sRPE, so almost half of what looked like noise was the training programme. On wellness, almost nothing: between 0.5% and 6%. That is why the comparison between matching days is the reference wherever you look at the week: the microcycle profile and the shadow on the charts (this MD-3 against your last three), the metrics measured only occasionally, and also the cards and the map that tell you who to watch. There, on metrics that add up, each training or match day is compared with its last 8 matching MDs and, until there are enough of them — around ten weeks —, with days of the same kind: training against training, match against match. A rest day never enters the norm of a training day, and the program states which of the two comparisons it used. The Flags rules that look at load and the single-metric flag in Reports use this comparison too, with thresholds corrected for how many matching days there are.

A match sits at values no training session reaches. On the simulated season match day sits, as a median, 1.05 standard deviations above a norm that mixes every day, and exceeds 1.5 in 17% of cases against 2.6% for the other days: with matches and training in the same pile, 53% of the alerts fell on 12% of the days.

That is why the norm can keep matches apart — the choice is yours, and the program always says how many matches the window holds — and match day — in the typical match, the match target, availability and the post-match sheet; the personal norm keeps it whole — is split: the part played is match, the extra work of those who stayed on the bench is training, as the consensus standard on football injuries requires. And 90′, 20′ and 0′ are three loads that cannot be compared: without minutes played the program writes “n/a”, never that a squad member played.

On a goalkeeper GPS says little: distances and high speeds are low because of the position, not because of condition. The 19.8 km/h threshold, which for an outfield player is 53–61% of top speed, is 84% for the goalkeeper of the simulated season. A traffic light built on GPS would call him “underloaded” all season, and would teach people to ignore his alerts.

That is why the goalkeeper never enters the position norm of outfield players: goalkeepers form a group of their own. And where a screen shows a GPS metric for him, the program writes that there it measures the position, not the condition — without hiding anything: for him force plates, wellness and sRPE matter more.

GPS and wellness are measured every day; force plates once or twice a week. A window expressed in days that works for load is almost empty for the jump: 28 days hold 4 or 5 tests, half of the 8 values needed to estimate a norm.

It is a mistake we made and measured: with the load window, the rule “high load plus falling jump” stayed unassessable across the whole squad — 11 athletes out of 11, even those with more than 40 tests. Every window has its own length, and where the source is too sparse the program declares it instead of computing anyway. Since the norm’s window is derived from that metric’s real cadence — the jump rule has a window of its own, 90 days, and where a norm does not yet exist the last test is compared with the previous one — 95% of the force-plate days that used to stay silent now say something.

A zero is a measurement: it means that on that day the athlete did not work. A gap is the absence of a measurement, and it can mean they were away with the national team, that the GPS did not start, or that nobody uploaded the file. Confusing the two distorts ACWR, monotony and z-score invisibly: a missing load read as a zero lights up a false “below their norm” on that day — on the simulated season it would have happened 10 times — and then widens the variability of the norm, so the truly unusual days stand out less.

The calendar decides: a day marked as rest with no measurements counts as zero, a training or match day with no measurements stays a declared gap. And on the charts an absence can be seen: when the missing days exceed three times the series’ usual step, the line breaks, and the empty space is as wide as the absence, up to 40 days.

Re-analysing the same training files with a new version of the manufacturer’s software, the accelerations counted above 1.5 m/s² fell on average from 251 to 177 per session, without the athletes changing anything: the way of counting changed. The same happens when you change vendor, and a personal norm built across the change mixes two instruments.

That is why the program marks the day of the change with a line on every chart, and when a calculation window crosses it, it says so next to the number. The vendor and the software version are declared at import: a day without provenance is not read as a change.

These are two different questions and they have to be kept apart. The first belongs to the instrument: from how much each athlete’s test changes from one session to the next — the differences between consecutive tests, pooled across the whole squad — you get the typical error, and from there the smallest detectable change. Below that number, two measurements are not different: they are the same measurement repeated. The repeated trials within a session — the three trials of a CMJ are separate rows, not an average already taken — answer another question: how precise the day’s number is.

The second is a decision, not a property: how much it has to change to be worth looking at. The program proposes it from your own data, in two ways that answer two different questions, and leaves the declaring to you. That the two numbers diverge is information, not a fault: if the noise is larger than the threshold that matters, that instrument cannot see what would interest you.

A z-score reads as if the mean and standard deviation were known. They are not: they are estimated from the athlete’s history, and on a metric tested every 7 days that history is a dozen values. A fixed threshold applied to such an estimate is more confident than it has any right to be.

So the thresholds the program chooses — traffic light, cards, map, overviews — are corrected for how many values the window really holds: the right distribution is a Student’s t, not a normal. With 28 values the attention threshold moves from 1.5 to about 1.57 deviations; on load compared with 8 matching MDs, to about 1.80; on the force plates, with 12 values, to 1.69. The thresholds you write in Flags — load, wellness, jump and asymmetry — are corrected the same way too, for how many values that athlete’s norm has. The correction assumes roughly bell-shaped values: where they are strongly skewed it remains an order of magnitude, and the Z-score and Flag panels of the Reports say so when they measure strong skew.

A system that flags 40% of the days flags nothing: it teaches people to ignore the colour. It happened to us too — with a percentage-change threshold the alerts on total distance reached 48% of athlete-days, because inside a microcycle the hard day sits naturally 20% above its reference.

Moving to the z-score against the work norm — each day against their last days with the same MD, match against matches, divided by how much that athlete varies — the alerts fell to 11%, and the ratio between reds and oranges straightened out. Above the deviation of a single number there are four rules that look at two things together: high load with falling wellness, above-usual work since the previous test with a falling jump, a return from absence with load already high, asymmetry far from their own norm. Every alert carries the numbers that lit it up, and a rule that could not be assessed declares it: “I could not look” is not “everything is fine”.

A threshold that is the same for everyone always flags the same people. Some athletes live at 12% on a test and that is normal for them — a dominant side, an old injury long healed — while others go from 2 to 9% without crossing anything. With a fixed 10% the first lights up every week and the second never does: across the simulated season 40 of the 45 alerts fall on the same athlete.

The program reads both values of the same test, writes which side is weaker and with which formula — on a time, the weaker side is the slower one — and compares the asymmetry with the band of their usual: the 25th–75th percentile of their last 8 tests, earlier ones only. “Worth a look” when it leaves the band by at least the band’s own width: with the 75th alone a stable athlete would be out on one test in three, this way on one in ten (31% → 10% of tests). It works for any pair of metrics, from any provider, including ones entered by hand.

The ratio between acute and chronic load is the most debated number in the load literature, and the criticism is serious: computed the classic way, the acute week sits inside the chronic window, and the two numbers correlate by mathematical construction, not because of what the athlete does.

That is why the default is uncoupled — the chronic load is the 28-day exponential average read as it was seven days earlier, so the acute week does not enter it — and exponentially weighted. Absences are excluded from the chronic load instead of counting as zero: counted as zero, a return with a perfectly normal load would come out at 3.0. The 0.8–1.3 band is drawn in grey: it is a reference from the literature, not a safe zone. And the ACWR does not enter the readiness score: it is shown alongside, as a descriptor.

There is no risk score, no injury probability, no recommendation on fitness to play. This is not commercial caution: a system that predicts is a clinical decision support tool, with everything that entails — and prediction promises, in this field, have not held.

What the program says is: this value sits this far from this athlete’s norm, and here is the raw data behind it. The vocabulary is chosen on purpose — “outside their norm”, “falling”, “worth a look” — and it does not contain “at risk”. Interpretation stays with the qualified practitioner, and the sentence sits inside the product, on the screens where the thresholds are configured.

A CSV file from any vendor, and the squad: no need to change vendor or install a connector. The columns are mapped once and the profile is saved; the season dates and the fixtures in the calendar are what the microcycle needs, and they go in over one afternoon.

You do not need to decide anything before seeing the first number: the defaults are sensible, declared and all adjustable, and the guided first steps walk you through the setup one stage at a time.