Forecast accuracy by horizon and the lag your money commits at
Nov 20, 2026, 9 min read

An accuracy figure quoted without a lag is undefined, so forecast accuracy by horizon should be reported at the lag your money commits at. Most published accuracy is lag one, which scores a forecast made after every commitment for that period had already been placed. Aggregation moves the number as far as the lag does, and it is the easier of the two to pull by accident. Plot error at every lag out to your longest commitment; the curve rises and then flattens, and where it flattens tells you which horizon is worth working on.
The slide says forecast accuracy is 84 percent, up two points on last year. The buyer sitting in the same meeting places factory orders fourteen weeks before the goods sell, and nothing in the pack tells her how good the forecast was fourteen weeks out. She has never asked, because nobody has ever offered the number, and the 84 percent has an air of authority that makes the question feel awkward.
The 84 percent is almost certainly the error between last month's forecast and last month's actual. It describes a forecast made at a point when every commitment for that month had already been made. It is a measurement of something no decision depended on.
A forecast has two dates on it
Every forecast value has an origin, which is when it was made, and a target, which is the period it describes. The gap between them is the lag. An accuracy figure quoted without a lag is undefined, in the same way that a lead time quoted without a unit is undefined, and yet accuracy travels around most businesses with no lag attached at all.
The lag that matters is set by the decision, not by the reporting calendar. If the purchase order goes out fourteen weeks before the sale, the forecast that mattered was the one standing fourteen weeks out. Everything the demand planner did in the intervening thirteen weeks improved a number that could no longer change anything except the expedite decision.
So the first thing to establish is your own commit lag, per decision type. Long lead time raw material might commit at twenty weeks. Domestic finished goods might commit at three. Promotional volume might commit at eight because of the retailer's booking calendar. Capacity might commit at a quarter. Those are the lags to measure at, and there will be more than one of them.
One item, two lags
Take one item over six months. Actuals were 1,180, 1,340, 1,010, 1,620, 1,450 and 1,100, summing to 7,700 units.
The lag one forecasts were 1,050, 1,225, 1,120, 1,430, 1,330 and 1,235. Absolute errors of 130, 115, 110, 190, 120 and 135 sum to 800. Divide by 7,700 and the weighted absolute percentage error is 10.4 percent, or 89.6 percent accuracy.
The lag four forecasts, made when the orders were actually placed, were 940, 1,020, 1,240, 1,150, 1,190 and 1,380. Absolute errors of 240, 320, 230, 470, 260 and 280 sum to 1,800. That is 23.4 percent error, or 76.6 percent accuracy.
Same item, same six months, same actuals. Thirteen points of difference, and only one of the two numbers governed a purchase order.
The buffer consequence is larger than the accuracy gap suggests, because safety stock scales with the standard deviation of the error rather than with the mean error. The lag one residuals have a standard deviation of about 138 units. The lag four residuals have a standard deviation of about 309. A buffer sized on lag one error is roughly 45 percent of the buffer the commit lag actually requires, which is a service problem that no amount of tuning the service factor will find, because the input was wrong rather than the formula.
Worth noting in passing what else falls out of that arithmetic. The mean lag four error is plus 130 units a month, meaning the forecast ran systematically low by about 11 percent of average demand. Bias at the commit lag is a separate measure with a separate cure, and BB1 has it.
Same forecasts, different aggregation
The second lever moves the number as far as the lag does, and it is easier to pull by accident.
Take three items in one family in June, all at lag four. Item A sold 1,100 against a forecast of 1,380, so the forecast was 280 units high. Item B sold 640 against 500, so 140 low. Item C sold 910 against 780, so 130 low.
At item level the absolute errors sum to 550 against 2,650 units of actual, which is 20.8 percent error, or 79.2 percent accuracy.
At family level the forecast was 2,660 against an actual of 2,650. Ten units out. That is 0.4 percent error, or 99.6 percent accuracy.
Nothing changed except the level at which the subtraction happened. Errors of opposite sign cancel when you aggregate, and the more items you aggregate, the more they cancel. A company reporting accuracy at family or category level is reporting the cancellation, and the plant that builds item A built 280 units of the wrong thing regardless.
Time buckets do the same thing along the other axis. The same three items measured in weekly buckets for that month came out at 34 percent error, because timing errors inside the month cancel when you sum to a month. Report at the bucket and level at which the decision is made: if the schedule is weekly at item-location, weekly item-location accuracy is the number, however unflattering.
There is a third version choice that quietly rescues bad numbers. If your accuracy report reads the forecast currently sitting in the system for a past period, it may be reading a value updated after the fact. Snapshot the forecast at each origin, store it, and compare the snapshot. Any accuracy process that cannot show you a frozen lag four snapshot from six months ago is measuring something it cannot define.
The horizon curve, and where it flattens
Compute the error at every lag from one out to your longest commitment, and plot it. For the item above the curve runs roughly 10.4 percent at lag one, 14.8 at lag two, 19.6 at lag three, 23.4 at lag four, 29.1 at lag six and 33.0 at lag nine.
Two features of that curve are worth more than the individual points.
The first is where it flattens. Between lag six and lag nine the error moves 3.9 points, against 13 points between lag one and lag four. Past the flattening point the model has run out of information and is essentially reproducing the seasonal level, so a better model buys almost nothing there. If your commit lag sits beyond the flat section, the lever that improves your inventory is a shorter lead time rather than a better forecast, and that changes what you should be spending money on.
The second is the comparison against a naive benchmark at every lag. Run seasonal naive alongside, and for this item it gives 19.8 percent at lag one, 21.4 at two, 23.2 at three, 25.0 at four, 30.5 at six and 33.4 at nine. The model's advantage is 9.4 points at lag one, 6.6 at two, 3.6 at three, 1.6 at four, 1.4 at six and 0.4 at nine.
Building the curve needs one table and no new software. Store every forecast as a row with four fields: the origin period, the target period, the item-location key and the value. Lag falls out as the difference between the two periods, actuals join on the target period, and the curve is a group-by. The storage is trivial next to what a planning system already holds, and the discipline is to write the row at the moment the forecast is published rather than reconstructing it later.
Use rolling origins so every lag gets as many observations as the history allows. With 36 months of stored forecasts, lag one gives you 35 origin and target pairs and lag nine gives you 27, and each lag should be reported with its observation count so that nobody reads the long end with more confidence than it deserves.
That is the finding that changes a conversation. The model earns its keep at short horizons, and the business commits at lag four, where it is 1.6 points better than repeating last year. Scaling error against a naive benchmark properly is BB2's subject, and attributing the remaining gain to the people and steps that touch the forecast is D7's; the horizon version of the question is simply whether the advantage still exists where the money is spent.
What lag one accuracy invites
Measure accuracy at lag one and three behaviours follow, none of which requires anyone to act in bad faith.
Planners refresh late. Publishing the monthly forecast on working day eight rather than working day two, using the first week of actuals, is an easy few points of lag one accuracy and changes no decision.
Reports drift upward through the hierarchy. When accuracy at item-location looks poor, the report gets rebuilt at item level, then at family. Each step is defensible on its own and the cumulative effect is a number that no longer describes anything anyone plans.
Planners work the wrong queue. Ranked by percentage error, the top of the list is always the smallest items, because an item selling four units forecast at six is 50 percent wrong while an item selling 40,000 forecast at 41,600 is 4 percent wrong and 1,600 units out. Rank the queue by absolute error in units or in margin, and the list changes completely.
A fourth habit worth naming: reporting accuracy as one hundred minus the percentage error, floored at zero. A month where the forecast was double the actual produces an error above 100 percent, gets floored, and enters the average as zero rather than as the disaster it was. Report the error rather than the accuracy, or at least keep the unfloored figure visible.
What to pair it with
Accuracy at the commit lag, on its own, is still gameable by anyone who can choose the aggregation. Three companions close most of it.
Bias at the same lag, because a forecast can be accurate in absolute terms and consistently low, which is the version that empties shelves, or consistently high, which is the version that fills warehouses.
A scaled benchmark at the same lag, so that a hard item scoring 40 percent error and an easy item scoring 12 can be compared on whether the model beat the obvious alternative.
The commit lag distribution itself, which almost nobody measures. Pull last year's purchase orders and compute the gap between order date and consumption date. If the median is nine weeks and the tail runs to twenty, you have two lags to report, and the tail is where the cash sits.
Where this stops
The long end of the curve is noisy, and the noise is structural. With 24 months of history, lag nine gives you about 15 origin and target pairs, and a three point difference across 15 observations is well inside what sampling variation produces. Report the long lags with the number of observations next to them, and resist tuning anything on the basis of a small improvement out there.
Percentage error also breaks where half of most catalogues live. An item with zero demand in a period gives a division by zero, and the usual workarounds either drop those periods, which flatters the result, or substitute a small number, which distorts it. Intermittent items need a different instrument, and D5 covers the forecasting side while BB2 covers the scaling.
The curve also tells you the size of the problem rather than whether the problem is solvable. Some series carry no usable signal at the horizon you need, and a horizon curve that sits on top of the naive benchmark from lag two onward is telling you that; D11 has the argument about which series those are and what to do instead of buying a better model.
Identify the lag at which your largest cash commitment is made, rebuild last year's error at exactly that lag, and put the two numbers side by side in the next review.
More on this in metrics on the DemandLabs platform.




