What Accuracy Is Worth
Accuracy has a price and a value, and they are rarely compared. How to work out what a percentage point of error costs, and when more modelling stops paying.
Accuracy is worth whatever the decision it feeds costs when it goes wrong, and that figure is almost never calculated. Without it, every argument about forecasting effort is an argument about taste, and the usual outcome is that effort goes wherever the loudest complaint came from.
The error measure in “What Accuracy Is Worth” needs both the original forecast and a trustworthy record of what actually happened. Teams researching attendance point system can use view the supporting platform page to strengthen the actual-work side of that comparison without treating activity data as proof of demand, performance or intent.
The calculation is not difficult. It takes an afternoon and it settles several years of argument.
The assumptions behind “What Accuracy Is Worth” can also be tested against Tableau analytics resources. Record which guidance applies to the available series and which does not.
The two sides of being wrong
Forecast too low and the cost is overtime, agency cover, queues, missed service levels, and the slower damage of people leaving. Forecast too high and the cost is paid idle time, and occasionally a redundancy process.
- Under: overtime premium, agency cover, queue, missed service, and the attrition that follows a run of short weeks.
- Over: paid idle time at standard rates, absorbed into other work, and occasionally a redundancy process.
These are rarely symmetric, and which one is dearer decides how the forecast should be biased on purpose.
Pricing a point of error
Take the decision. Work out what a 1% error in the forecast changes in the resource it drives. Price that at the marginal rate — overtime rate, agency rate, idle hourly cost — not at average salary.
For a team of fifty, 1% is half a person, which over a quarter is roughly 280 hours. At an overtime premium that is a number anybody can recognise.
The asymmetry is the finding
Most operations find that being under is three to five times dearer than being over, because the under side is paid at premium rates and carries service consequences, and the over side is paid at standard rates and absorbs into other work.
That asymmetry, once known, says the forecast should sit deliberately above the central estimate for staffing decisions — and it says by how much.
Deliberate bias is not dishonesty
As long as it is named. "Staffing is planned at the 70th percentile of the forecast, because being short costs about four times being long" is a stated policy that anybody can question.
What is dishonest is the same margin applied invisibly inside the forecast, which makes the forecast unassessable and the policy undiscussable.
Where the margin should live
In the staffing rule, not in the forecast. The forecast says what is expected; the rule says what is resourced given the asymmetry.
Two numbers, two owners, and the ability to change the risk appetite without rewriting the model.
When more accuracy stops paying
When the cost of the remaining error is smaller than the cost of reducing it. A model that takes two days a month to maintain and removes one point of error from a decision where a point is worth £400 a quarter is not worth running.
This calculation retires more forecasting projects than any other, and it should be run before the project rather than after.
The error you cannot remove
Some of it is irreducible: genuine randomness in arrivals, events nobody could know about. A good method approaches that floor and cannot go below it.
Knowing roughly where the floor is — usually estimable from the variation between comparable days — stops people demanding accuracy the world does not contain.
Accuracy at the wrong horizon is worth nothing
A week-ahead forecast improved from 4% to 2% error, where no decision is taken at a week, has produced nothing. The same effort at the quarter, where recruitment is decided, might be worth a great deal.
Effort should follow the decision value, and it almost always follows the data availability instead, because the near horizon is where the data is good.
The figure to publish
One line: what a point of error is worth on each side, for each decision. It belongs next to the error record, and it turns a discussion about whether 6% is good into a discussion about whether 6% is affordable.
The second discussion can be concluded.
Who should own the number
Whoever holds the budget that absorbs the cost, not the planner. The planner can do the arithmetic; the valuation is a management statement about risk appetite and it belongs to a manager.
A planner who sets it alone has quietly set the organisation's tolerance for being short-staffed.
Reviewing it
Annually, or when rates change. Overtime premiums move, agency rates move sharply, and the asymmetry can invert — in a labour market where cover is unobtainable at any price, being under stops being expensive and starts being impossible.
That is a different regime and it wants a different margin, which is why the figure is reviewed rather than set once.
The cost that falls on people
Not all of it is money. Chronic under-forecasting is paid in cancelled leave, in short-notice overtime and in the attrition that follows, and those costs are real and slow.
They belong in the valuation even though they resist pricing, and naming them in the same line as the financial cost is usually enough to change how the asymmetry is set.
What it changes
It converts forecasting from a craft argued about on aesthetic grounds into an investment with a return anybody can check, and it is usually the first thing that gets a planning function the resource it has been asking for.