Why chasing forecast accuracy is the wrong goal
The target everyone agrees on, and almost no one questions
Ask a planning team what they are trying to do and “improve forecast accuracy” comes back almost reflexively. It sits on scorecards, in objectives, on the wall. It feels unarguable: a more accurate forecast is a better forecast, so chase the number down and the operation gets better. The trouble is that accuracy is a means, not an end. The actual goal is a plan that delivers the right service at the right cost — and beyond a certain point, a more accurate forecast barely moves that, while a great deal of effort is spent shaving decimal places off a metric that was never the thing that mattered.
This is not an argument for sloppy forecasting. It is an argument for knowing what accuracy is for, measuring the part of it that actually drives decisions, and spending the saved energy on the thing that determines whether a plan succeeds: how well it survives the error that will always remain.
Error is not optional, and you already know its floor
No forecast of human behaviour is exact. Even a perfect model of demand runs into the fact that arrivals are random: the same Tuesday, lived a hundred times, gives a hundred different contact counts. That irreducible scatter — the Poisson floor — sets a hard limit on how accurate any forecast can be. Chasing accuracy past that floor is chasing noise, and noise does not respond to better models, more history, or longer planning meetings. The first discipline is recognising where the floor is, so you stop spending effort below it.
Once you accept that error is permanent, the question changes. Not “how do I eliminate the error?” but “given that error remains, is my plan robust to it?” That is a completely different — and far more useful — place to put your attention.
Not all error costs the same, which is what the headline metric hides
The standard accuracy metric is MAPE — the average percentage you were off, period by period. It is easy to quote and quietly misleading. It treats every miss as equal: a 10% error on a dead Sunday counts the same as a 10% error on your busiest Monday, even though only one of them sends you short of staff when it matters. And it says nothing about direction. A forecast that is 6% high every single week and one that is 6% high one week and 6% low the next have the same MAPE, but they are not the same forecast at all. The first is quietly costing you money on overstaffing every week; the second is roughly self-cancelling.
That directional error — bias — is the one that compounds, and it is the one MAPE buries. A forecast with low bias and moderate noise is far more useful than one with low noise and a persistent lean, because bias mis-sizes the whole operation in one direction, week after week, while noise averages out across a roster. If you measure one thing about your forecast, measure bias. If you measure two, add WAPE — the volume-weighted error — because it counts your busy periods for what they are worth instead of letting quiet days flatter the average. The forecast accuracy calculator gives you bias, WAPE and the tracking signal alongside MAPE precisely so you stop quoting the weakest of the four.
Which input is wrong matters more than how wrong the forecast is
Even a flawless contact forecast is only half the staffing equation. The number of people you need is driven just as hard by average handle time, by shrinkage, and by the service target you are planning to — and an error in any of those moves the roster too. Spend all your effort on the contact forecast and you can still be badly staffed because handle time came in heavier than planned or shrinkage ran above the policy figure you used.
This is where the goal really should sit: not “how accurate is the contact forecast?” but “which of my inputs is the plan most exposed to, and have I buffered against that one?” Volume and AHT usually have roughly equal leverage, because both feed straight into the workload. Shrinkage is often the quiet third lever and the one most likely to be set from a stale assumption rather than measured. The service-level target, the thing people argue about most, typically moves staffing the least. Knowing that order — for your operation, not in general — is worth more than another point of forecast accuracy, because it tells you where to put your contingency and where to stop worrying.
A useful forecast is one you can act on, not one that is merely close
The reframe is simple to state and surprisingly hard to live by: the job of a forecast is to produce good decisions, not accurate numbers. A forecast that is 3% more accurate but arrives too late, in a form no one can roster to, or without any sense of its own uncertainty, is worse than a slightly looser one that is timely, decision-ready and honest about its range.
Three things make a forecast actionable, and none of them is accuracy. It has to be timely — available while there is still time to hire, train or move shifts. It has to be at the right grain — an annual total is useless for a roster, which lives or dies on the intraday profile, while an interval-level forecast is the wrong tool for a budget. And it has to carry its uncertainty — a range, not a single number, so the people downstream can plan for the spread instead of being ambushed by it. That last point is the bridge to forecasting with ranges, which is the practical form of admitting that the error is real and planning around it anyway.
What to do on Monday
Replace “improve forecast accuracy” on the scorecard with two sharper targets: drive bias towards zero, and keep WAPE inside a band you have agreed is good enough for the decisions it feeds. Stop reporting MAPE as the headline; report bias first. Run a sensitivity check on your biggest planning lines so you know whether volume, AHT or shrinkage is your real exposure, and size your buffer against that input rather than spreading worry evenly. Forecast material numbers as ranges, and track how often the actual lands inside the band — a healthy range is right about three times in four, not always. And when accuracy is already past its useful floor, have the confidence to stop polishing it and put the time into the plan instead.
None of this lowers the bar. It moves it to where it belongs. The best planners are not the ones with the lowest MAPE; they are the ones whose plans hold up when the forecast is wrong — because they always assumed it would be.
Conclusion
Accuracy is worth having, right up to the point where it stops changing decisions — and that point arrives far sooner than most scorecards assume. Measure the error that matters, which is bias before noise and weighted error before the simple average. Know which input your plan is most exposed to. Deliver the forecast in a form people can act on, with its uncertainty attached. Do that, and you will spend less time defending decimal places and more time building plans that survive contact with reality. That is the whole job.
Put the ideas to work: the forecast accuracy calculator (bias, WAPE and tracking signal), the staffing sensitivity tool, the intraday profile builder and the FTE & budget builder. Pair with forecasting with ranges and Poisson and natural noise.