Missing training data is never missing at random

The sessions that go unrecorded are the ones with something to say, so filling the gaps with averages buries it.

Look at where the holes are in a season of load data. They are not scattered evenly, and that is the whole problem.

Units do not get worn on the days when the schedule is chaotic. Nobody collects wellness forms on the morning after a night flight. Sessions on an away pitch, in a borrowed facility, at a tournament, in the rain when the tablet stayed inside: those are the ones that go unrecorded, and they are the same sessions that were unusual in ways worth knowing about. The gaps in the file mark the days when the routine broke, which is exactly when the athletes were doing something outside the routine.

Then the analysis has to do something with the holes, and the standard moves make it worse. Drop the missing days, and every rolling average is computed over the days that went normally, so the season looks tidier than it was. Fill them with each player's mean, and you have inserted a typical day into a slot that was atypical by definition, which shrinks the variance and makes a disrupted week look like a routine one. Carry the last value forward and you assert that nothing changed on the day when most things did.

All three of those are decisions about what the missing days contained, and none of them are usually described as decisions. They are defaults in a piece of software.

The pattern of missingness also differs by player, which is the part that makes cross-squad comparison unsafe. The senior professional with a private routine records less than the twenty-year-old who does everything with the group. Compare their totals and you are partly comparing their compliance with the recording process. The player with more complete data will appear to be training more.

What I would do is keep a column that counts the holes. Sessions recorded out of sessions held, per player, per month. Publish it next to the load numbers. It costs a formula and it tells any reader how much weight the row can carry.

And when somebody presents a beautiful season-long trend, the first question is not what caused it.

The first question is which weeks are missing, and whether they went missing for reasons connected to the very thing being plotted.