Improving the wrong number
Forecasting programmes are judged on error against the mean. It is a clean metric and it improves steadily, which makes it satisfying to work on. It also has surprisingly little to do with what a bad week costs.
Operations are asymmetric. Being over by ten per cent costs idle hours; being under by ten per cent costs missed commitments, overtime and the reputational tail that follows. A model tuned to minimise average error will happily trade a cheap mistake for an expensive one.
The plan does not need to know what will happen. It needs to be affordable across everything that might.
Plan the spread
- 01State the range, not the point. A single number hides exactly the information a planner needs.
- 02Price both directions of error, because they are not equal and never were.
- 03Buy flexibility where the spread is widest, cross-training, standby agreements, deferrable work.
- 04Re-plan on a cadence the operation can actually act on. A perfect weekly forecast is useless to a team that commits daily.
Flexibility beats precision
A team that can move people between two functions in an hour absorbs more variance than a model improvement of several percentage points. The first is a scheduling and training decision; the second is a data science programme. They are usually funded in the wrong order.
This is not an argument against forecasting. It is an argument for being honest about what a forecast is for, narrowing the range of plans worth preparing, not choosing one and hoping.
