Your Own Turnover Data
Everything else on this site is somebody's research. This is the part that is about you, costs nothing, and uses data already sitting in a payroll system. For a related reference, see Chinese overtime calculations.
The four cuts
By manager. The most informative and the least performed. Departures per head reporting to each manager, over two or three years.
By tenure at departure. People leaving at four months and people leaving at four years are two different problems: the first is hiring or onboarding, the second is trajectory.
By date, against events. Plot departures monthly and mark policy changes, reorganisations, promotion rounds. Clustering is the finding.
And by regretted against not. Somebody has to decide which departures you minded, and that judgement is the point — an undifferentiated turnover rate treats a good exit and a bad one identically.
What each cut tells you
Manager clustering points at the largest controllable factor, and it points at it specifically rather than in general.
Early tenure departures mean the job was described differently from how it is, or the first weeks were bad. Both are fixable and cheap.
Late tenure departures mean nobody discussed where this goes.
Event clustering attributes a cost to a decision that was made without one. A mandate carrying roughly 14% additional attrition is a number somebody should have seen beforehand.
The trap in small numbers
Below roughly thirty departures, most differences are noise.
Three departures under one manager of eight looks alarming and may be chance. Two years of data helps; one does not.
What survives small numbers is the pattern rather than the rate: the same manager across two years, the same tenure band repeatedly, clustering after a specific date.
Say "worth a conversation" rather than "the cause", because the data at this scale supports the first and not the second.
What to do with a manager who clusters
Not dismissal. Most managers were promoted for the previous job and given no training, and replacing one with another untrained promotion repeats it.
A conversation with the departures, if any will speak. Better a month after leaving than on the way out.
And a look at what that manager was given — span of control, an impossible team, a role nobody could do well.
Where to keep it
A sheet, updated when somebody leaves. Date, role, tenure, manager, regretted or not, stated reason.
Six fields, thirty seconds per departure, and in two years it answers questions no purchased system will, because it contains the judgement column that no system can populate.
The short version
- Four cuts of data you already hold: by manager, by tenure at departure, by date against events, and by whether you minded
- Manager clustering points at the largest controllable factor specifically; early tenure means hiring or onboarding, late tenure means trajectory
- Event clustering attaches a cost to a decision made without one
- Below roughly thirty departures most differences are noise, and what survives is the pattern rather than the rate
- Say "worth a conversation" rather than "the cause", because the data at this scale supports only the first
- Six fields in a sheet, thirty seconds per departure, and the judgement column is the one no purchased system can fill
For additional context on this topic, see PetaPixel.