Six weeks to a forecast that outperforms yours See how
Back to planning notes

Your bias number is near zero and the warehouse still filled up

Nov 12, 2026, 9 min read

In short

Forecast bias is the average of the errors, so it can read near zero across a catalogue while individual items run consistently high, because opposite biases cancel on the way up. The sign convention is rarely written down, and actual minus forecast and forecast minus actual give opposite readings of the same quarter. Bias accumulates where random error cancels, so an item selling 1,000 units a week against a forecast running 4 percent high adds 520 units over a quarter, about half a week of cover. Measure it as a running total at item level rather than as a period average across the portfolio, and pair it with a magnitude metric so that a flat bias number cannot pass unexamined.

The quarterly pack has two numbers on the forecast slide. Accuracy at 78 percent, flat on last quarter. Bias at 0.2 percent, with a footnote saying it sits inside a tolerance of plus or minus two. Nobody asks about either one, because both look fine.

Nine slides later there is a chart of weeks of cover across the network, and it has moved from 5.1 to 6.4 over the same period. The commentary attributes it to phasing.

Both numbers were computed correctly. The 0.2 percent is near zero because of three choices somebody made when the report was built, and each choice was reasonable in isolation. Different choices on the same data give you a bias of minus 9.7 percent, and that version would have predicted the cover chart.

The sign convention that never gets written down

Bias is the average of the errors, and error has two conventions. Actual minus forecast makes a positive number mean you sold more than you said. Forecast minus actual makes a positive number mean you forecast more than you sold. Both appear in commercial planning systems, sometimes in different modules of the same one.

The consequence shows up when two teams report to the same meeting. Demand planning reports plus 3 percent from the planning system, supply reports minus 3 percent from a warehouse report, and half an hour goes into establishing that they agree.

Write the direction on the axis in words. Positive means we shipped more than we said. It takes six words and it removes an argument that otherwise recurs monthly.

Four defensible numbers from one quarter

Here is a quarter of data for four items, using actual minus forecast, where a positive error means the forecast came in low.

Item A: actual 10,000, forecast 9,000, error plus 1,000, percentage error plus 10 percent. Item B: actual 200, forecast 400, error minus 200, percentage error minus 100 percent. Item C: actual 50, forecast 20, error plus 30, percentage error plus 60 percent. Item D: actual 9,000, forecast 9,800, error minus 800, percentage error minus 8.9 percent.

Total actual is 19,250 and total forecast is 19,220, so the sum of errors is plus 30 units. Divide by total actual and bias is plus 0.16 percent. That is the ratio of sums, the version most planning systems report by default, and on this data it says the forecast is very slightly conservative.

Take the unweighted mean of the four percentage errors instead. Ten, minus 100, 60 and minus 8.9 average to minus 9.7 percent, which says the business systematically over-forecasts by about a tenth. Same four items, same quarter, opposite conclusion and a different order of magnitude.

Count the signs and you get a third answer. Two items over-forecast, two under-forecast, which reads as balanced.

None of these is a mistake. The ratio of sums is dominated by the two large items, so it describes the aggregate exposure and says nothing about the catalogue. The mean of percentage errors gives every item one vote, so item B, which sells 200 units, carries the same weight as item A, which sells 10,000. Percentage error is also bounded at minus 100 percent when you over-forecast a small item to zero sales and unbounded on the other side, which drags the unweighted mean downward on any catalogue with a long tail.

There is a fourth cut, and in a seasonal business it is the one that matters. Sum the errors by week rather than by item. A forecast that runs 6 percent high for the first six weeks of a quarter and 6 percent low for the last seven averages to roughly zero, and the stock that arrived during the first six weeks is still sitting there. The mean of the weekly bias figures is near nought while the mean of their absolute values is 6.

Bias is a running total, so measure it like one

The reason a small bias fills a warehouse is that it accumulates while a random error cancels. An item selling 1,000 units a week with a forecast that runs 4 percent high adds 40 units of unconsumed receipt every week. Over a quarter that is 520 units, half a week of cover on that item. Over a year it is two weeks.

What happens next depends on the replenishment policy, and it is worth being precise because the two cases behave differently. In a reorder point system the bias raises the target and the stock settles at a new, higher level, so the damage is a step change in working capital rather than unbounded growth. In a plan-driven system that pushes to a forecast, distribution requirements planning being the common case, the error keeps accumulating until a human notices, which is why bias shows up first as a rebalancing problem at the branches.

The instrument for a running total is a tracking signal: the cumulative sum of errors divided by the mean absolute deviation, both computed over the same window. Dividing by MAD makes the number comparable across items of any size. Trigg proposed the smoothed version in Operational Research Quarterly in 1964, and Gardner set out the control limits and the practical failure cases in the Journal of Forecasting in 1983.

Take the item above. Weekly demand around 1,000, MAD of 120, forecast running 40 units high each week. After eight weeks the cumulative error is minus 320 and the tracking signal is minus 2.7. After thirteen weeks it is minus 520 and the signal is minus 4.3, which trips the limit of four that most systems ship with. The individual weekly errors never looked unusual, because 40 units against a MAD of 120 is inside the noise every single week.

Two implementation details decide whether this works. The MAD has to be estimated on a window long enough to be stable, since a smoothed MAD that adapts quickly will grow along with the bias and mask it. And the cumulative sum needs a reset rule, because a signal that has tripped and been acted on carries history that will trip it again.

What a bias target does to the person carrying it

Put a tolerance of plus or minus 2 percent on a planner's aggregate bias and you get three behaviours, all rational.

The first is sign chasing. A month that came in high gets corrected downward, the correction overshoots, and the series alternates. Average bias lands near zero and the variance of the error goes up, which costs more than the bias did.

The second is offsetting. An aggregate tolerance is met by having errors in both directions across the catalogue, so nothing forces the item-level bias down. The measured number improves while the cover chart does not move, which is the situation in the opening.

The third is asymmetry in what gets corrected. Over-forecasting shows up as inventory in someone else's budget, three months later, described as a phasing issue. Under-forecasting shows up as a service failure this week with a name attached. Fildes, Goodwin, Lawrence and Nikolopoulos examined judgemental adjustments at four companies in the International Journal of Forecasting in 2009 and found that downward adjustments improved accuracy considerably more reliably than upward ones, which is what you would expect if the two error directions carry different social costs.

The fix for the target is to measure bias where it does damage. Report it at item-location, count the items whose error has held the same sign for three consecutive periods, and weight that count by inventory value. A planner cannot offset their way out of that number.

What to pair it with

Bias on its own is easy to satisfy and hard to interpret, so it needs company.

A magnitude measure computed on the same population. A forecast can be unbiased and useless. Pair the directional number with a scaled error ratio, which BB2 covers, and insist that both are computed at the same grain, on the same periods, and on the same version of the forecast.

The cover trend by segment. Bias predicts an inventory consequence, so put the two charts on the same page. If bias is near zero and cover is climbing, one of them is measured on a population that excludes the problem.

Bias measured separately on the demand plan and the constrained plan. Most businesses have at least two numbers in play, and if you measure error against the constrained supply plan you are scoring your own allocation decisions rather than your read of demand. The consensus process that produces the second number has its own arithmetic, which B2 deals with.

A same-sign run counter. Three consecutive periods in one direction is a signal that a control chart on individual errors will not raise for another quarter.

When a biased forecast is the right answer

There is a case where the bias number is correctly non-zero, and it comes up more often than the reporting allows for.

If you plan to a service level rather than to an expectation, the quantity you commit is a quantile of the demand distribution and not its mean. An item planned at the 90th percentile will show a persistent negative bias by construction, because you meant to over-supply it. Measuring that plan against a zero-bias target and correcting it is a slow way of dismantling a service policy that somebody chose deliberately. D9 covers the mechanics of planning to a distribution.

Two related traps sit next to this one. A model trained to minimise absolute error is aiming at the conditional median rather than the mean, and for the right-skewed demand typical of retail counts the median sits below the mean, so the model will look under-biased even when it is doing exactly what its loss function asked. Kolassa made this argument for count data in the International Journal of Forecasting in 2016, and it applies to any slow-moving item scored on absolute error.

The other is censored history. If you stocked out, recorded sales are lower than the demand that existed, so the error you compute against those sales makes an accurate forecast look high. The measured bias is then partly a record of your own service failures, which is one reason to reconstruct the unconstrained series before scoring anything, as D1 sets out.

Where this stops

Bias tells you a direction and nothing about cause. A forecast can run high because the model is wrong, because a promotion was planned and cancelled, because a customer moved to a different channel, or because someone added a number to protect a supply commitment. The metric points at the series and the search still has to happen by hand.

It is also unstable on short windows. Thirteen weeks of a low-volume item gives you very few effective observations, and a tracking signal on that item will trip on noise. Aggregate slow movers into a family before scoring bias, and accept that item-level directional error on a series selling three units a week is not a measurement you can act on.

The one thing worth doing this week: recompute last quarter's bias as the ratio of sums and again as the unweighted mean of item percentage errors, and if the two disagree by more than a couple of points, the report on the wall is describing your two largest items rather than your business.

More on this in metrics on the DemandLabs platform.

See what DemandLabs could do for your plan

Request demo

More planning notes

View all

Nov 11, 2026

Service went up after go-live and so did four other things

The quarter three review has the chart on it. Line fill was 94.1 percent in the six months before go-live and 97.3 percent in the six months after. That is 3.2 points, roughly a million and a half of recovered margin on volume that would otherwise have been missed, and the programme is declared a success on the second slide.

Nov 10, 2026

Half your benchmark gap is a definition and a product mix

The slide has two bars on it. Yours at 2.1 inventory turns, a peer group median at 3.8, and a gap converted into something like eleven million of released cash. By the time the deck reaches the operating review, the number has been written into next year's plan and given an owner.

Nov 9, 2026

Bottleneck analysis that survives a change in the mix

The capital request is for a second filler at 3.4 million, and the justification is one line: the filler runs at 92 percent. It has been in the plan for two years. On-time-in-full has not moved in either of them.