Forecast accuracy can hide offsetting errors

A forecast can be right on average and wrong in every period.

That sentence sounds contradictory because forecast accuracy is often treated as one score. But a score is the end of an aggregation process. If positive and negative errors are added together before anyone looks at the rows, they can cancel. The resulting zero can be mathematically correct and operationally misleading.

The short answer is precise: a zero average signed error shows no net direction in the selected rows; it does not show that the forecast was close to the actual outcome. To understand a forecast, keep the signed error, its absolute magnitude, its distribution, its horizon, its denominator, and the decision it supports as separate objects.

What does sales forecasting mean before accuracy is summarized?

Sales forecasting is a time-bounded estimate of a future commercial outcome for a declared unit, population, information cutoff, and horizon. The unit may be opportunities, bookings, revenue, units, capacity, or another outcome. An open pipeline total can be an input, but it is not the forecast definition by itself.

A usable definition states the decision the forecast should support and the later actual that will evaluate it. Without those boundaries, a change in the number can reflect a changed population, stage rule, horizon, or denominator rather than a change in forecasting performance.

Field	Question before the score	What remains separate
Forecast object	What is being estimated, at which grain?	Pipeline inventory and expected outcome
Information cutoff	Which information was available when the forecast was issued?	Later explanations and hindsight
Horizon	Which period or event closes the forecast?	A different sales-cycle or reporting window
Actual outcome	Which later observation will be compared with the forecast?	A proxy or an open status
Decision	What action should the forecast inform?	Generic accuracy without a use

Why can sales forecasting fail even when the dashboard looks complete?

A dashboard can show a coverage multiple, a win rate, a mean absolute percentage error, and a bias score and still fail as a decision instrument. Here, “fail” means that the forecast cannot tell the team what changed or what to test. Use the following diagnostic cases before changing the model, widening the pipeline, or opening a new improvement workstream:

1. Pipeline inventory mistaken for the forecast object. If the dashboard total is an open opportunity sum, name the later actual outcome, horizon, and decision the forecast is supposed to support. Do not treat pipeline coverage as sales forecasting by itself.
2. Signed errors cancel before anyone sees the rows. If average signed error is near zero while periods are missed in opposite directions, keep signed error and absolute error visible at row level before the mean. Do not conclude that the forecast was close.
3. Bias collapsed into accuracy. If one score is used for both direction and magnitude, separate signed error from absolute error and state which decision each supports. Do not treat an unbiased mean as proof of a usable forecast.
4. Denominator or population shifted. If the score changes after a stage rule, segment filter, currency, or excluded-row change, freeze the denominator and population before comparing versions. Do not conclude that forecasting performance improved or regressed.
5. Horizon or information cutoff moved. If the comparison mixes issue dates, close periods, or hindsight fields, restate the cutoff and horizon and rebuild the rows. Do not audit a moving target as if it were one forecast.
6. Adjustment direction hidden. If overrides improve some rows and worsen others, keep direction, size, and final versus baseline error separate, as the demand-planning evidence requires. Do not average them into one “judgment helps” claim.
7. FVA answering a different question. If a process score is used to certify commercial accuracy, return to the row-level error objects this article defines. Do not let a process audit replace the forecast-versus-actual review.

The existing source bank supports these boundaries without supplying a universal sales-forecasting diagnostic. Fildes, Goodwin and De Baets keep accuracy and bias distinct in a demand-planning corpus. Fildes, Goodwin, Lawrence and Nikolopoulos show why adjustment direction and size need separate inspection. Neither study establishes a B2B sales-forecasting benchmark. Each diagnosis must name the forecast object, cutoff, horizon, actual, and decision it is evaluating.

Why can aggregate forecast accuracy hide offsetting stage errors?

Begin with a convention. In this article, the signed forecast error is:

forecast error = forecast - actual outcome

An underforecast is therefore negative and an overforecast is positive. Another convention can be valid, but the sign must be declared before the result is interpreted. The following four periods are synthetic. They are designed to expose the aggregation problem, not to represent a company or a market.

Period	Actual outcome	Forecast	Signed error	Absolute error
1	100	80	-20	20
2	100	120	+20	20
3	100	80	-20	20
4	100	120	+20	20
Mean	100	100	0	20

The signed errors sum to zero: -20 + 20 - 20 + 20 = 0. The mean signed error is therefore zero as well. If a team calls that result unbiased, the statement is limited and technically correct. The forecast had no net tendency to be too high or too low across these rows.

The absolute errors tell a different story. Every period is wrong by 20 units. The mean absolute error is 20, not zero. The table has no correct period to hide behind. The positive and negative signs describe direction. Removing the signs reveals magnitude.

That distinction is the first release gate for a forecast review. A directional summary can answer whether the process tends to overshoot or undershoot. It cannot answer how close each forecast was. Those are different questions.

One word called accuracy hides several measures

Forecast teams often use accuracy as a convenient label for a selected error measure. The label becomes dangerous when the measure is not named. The following objects should stay visible:

Object	What it answers	What it can hide
Mean signed error	Is the selected set directionally high or low?	Opposite errors and the size of each miss
Mean absolute error	How large is the typical miss in the selected unit?	Whether errors are high or low
Root mean squared error	How strongly do large misses affect the summary?	The reason for an extreme miss and the ordinary case
Median absolute error	What is the middle miss after magnitude is taken?	A small number of severe tail errors
Quantiles or tail shares	How is error distributed across cases?	The causal reason for the distribution
Error by horizon or segment	Where does the process behave differently?	The aggregate average across unlike objects

None of these is the single true measure. Each is a lens. The correct lens depends on the decision. A capacity decision may care about the upper tail. A staffing decision may care about persistent bias. A cash decision may care about the timing and size of misses. A forecast used to allocate scarce inventory may value an underforecast differently from an overforecast.

The word accuracy should therefore be followed by a definition: accuracy according to which error, over which horizon, for which object, with which denominator, and for which decision?

Bias is not the same as accuracy

Bias is directional. Under the convention above, a positive mean signed error indicates that forecasts are high on average; a negative mean indicates that they are low on average. An unbiased mean can be created by alternating errors, as the synthetic table shows.

Accuracy is about closeness under a chosen measure. A forecast can have low bias and poor accuracy if its positive and negative errors are large. It can have modest accuracy but noticeable bias if most errors are small and point in one direction. The two measures can move together, but they do not contain the same information.

This is not only a mathematical warning. It changes how a process should be reviewed. If bias is the only reported metric, a team may celebrate cancellation. If absolute error is the only reported metric, a team may miss a systematic direction that creates inventory, capacity, or cash exposure. The review needs both the sign and the magnitude before it chooses a response.

The denominator can move the result

Percentage error adds another choice. A common form divides the error by the actual outcome. That may be reasonable when the actual is a stable positive quantity. It becomes unstable when the actual approaches zero, equals zero, or differs in scale across items. A 20-unit miss on an actual of 100 is not the same percentage as a 20-unit miss on an actual of 1,000.

Other reviews divide by the forecast, an average of forecast and actual, a cohort total, or a revenue base. Each denominator creates a different comparison. The result may be useful, but it must be named.

The same problem appears when a forecast moves from units to revenue, from a weekly horizon to a quarter, or from a product-level object to an account-level aggregate. Addition is not neutral when the rows carry different prices, exposure, or timing. A total revenue forecast can look stable while a set of important items carries large offsetting errors underneath.

Before comparing two accuracy summaries, record:

1. the forecast object and unit;
2. the issue date and outcome horizon;
3. the sign convention;
4. the denominator, including zero handling;
5. the inclusion and exclusion rules;
6. the treatment of cancellations, substitutions, and missing actual outcomes.

Without that metadata, a change in the number may be a change in the measurement frame rather than a change in forecasting performance.

Distribution matters before the mean

The synthetic table has the same error magnitude in every period. Real error sets are usually less tidy. Consider two four-period sets under the same sign convention:

Set	Signed errors	Mean signed error	Mean absolute error	Shape
A	-20, +20, -20, +20	0	20	Repeated moderate miss
B	-40, 0, 0, +40	0	20	Two exact periods and two larger misses

Both sets have the same mean signed error and the same mean absolute error. They do not carry the same operating risk. Set A misses continuously. Set B is quiet twice and then misses more severely twice. A review that only reports the two means cannot distinguish the service, capacity, or escalation pattern.

The next check is therefore distribution. Inspect the median, upper quantiles, tail share, and the rows that generate the tails. Then split by horizon, product, segment, geography, or process stage when those categories are part of the forecast object. The split should be decided before looking for a preferred story. Otherwise segmentation becomes another way to select the number that sounds best.

What the empirical studies actually show

Fildes et al. (2025)¹ analyze 147,131 forecasts and actuals from 10 organizations and 22 business units. Their evidence is organized as six datasets of one-step-ahead system forecasts, final forecasts, and actual outcomes. The reported overall medians show 51.5% of SKUs with improved Forecast Value Added and 55.6% with improved bias. The result is useful for this article because the study treats accuracy and bias as distinct dimensions and reports variation across datasets and organizations.

It does not mean that 51.5% is a target for every forecasting process. It does not mean that bias is an adequate substitute for accuracy. It does not mean that a commercial team can transfer the result to a pipeline without testing the object, horizon, information set, and outcome definition.

The earlier Fildes et al. (2009)² field study analyzes 68,984 complete forecast triples across four companies. It reports that upward adjustments were wrong-signed 66% of the time in the manufacturers and 83% at the retailer, compared with 46% for downward adjustments. Those figures make adjustment direction worth recording. They do not turn a supply-chain result into a rule that every upward sales override is wrong or every downward override is useful.

Together, the studies support a disciplined measurement habit: preserve the baseline or prior forecast, the intervention if one exists, the actual outcome, and the error fields before a summary is used. The studies do not eliminate the need to declare the decision and the measurement boundary.

Do not let FVA answer a different question

The existing forecast value added process audit owns a different but adjacent object. It asks whether a defined intervention added value compared with a preserved baseline, using an actual outcome, a selected metric, and the cost of the review process. The existing forecast override article owns the judgmental forecast call and the accountability risk around moving a number.

This article begins before either conclusion. It asks whether the error rows were still visible when the process calculated its average. A final forecast can beat a baseline on mean absolute error and still hide a severe tail. A mean signed error can be zero for both baseline and final forecast while their error distributions differ materially. The FVA comparison is stronger when those underlying fields remain available.

That is the boundary between the routes. The FVA article is a process comparison. The override article is a judgment and governance problem. This article is an aggregation-order and error-distribution problem. They should link to each other, not compete to own the same phrase.

The row-level forecast-error audit

Before publishing a score, build a small audit table. It can be implemented in a spreadsheet, warehouse, or forecasting system, but the fields should remain inspectable.

Audit field	Required question	Failure if omitted
Forecast object	What exactly was forecast: units, revenue, opportunities, capacity, or something else?	Unlike rows are silently combined
Baseline and final forecast	Was an intervention applied, and what was the untouched value?	FVA and override effects cannot be separated
Actual outcome	What event closed the forecast window?	Error is calculated against a moving target
Signed error	Which direction convention is in force?	Bias changes sign or becomes uninterpretable
Absolute or squared error	How large was the miss regardless of direction?	Cancellation is mistaken for accuracy
Horizon	How far ahead was the forecast issued?	Short and long forecasts are treated as identical
Denominator	What base creates the percentage comparison?	Ratios change without a visible measurement change
Segment and tail flag	Which rows create persistent or severe errors?	Aggregate averages erase the operational problem
Decision cost	What does an over- or underforecast make the organization do?	Metric improvement is mistaken for decision value

Run the review in this order:

1. freeze the row-level forecast and actual definitions;
2. declare the sign convention and denominator;
3. calculate signed, absolute, and at least one distributional view;
4. inspect horizons and segments before the aggregate is presented;
5. retain excluded or missing rows with a reason code;
6. compare a baseline with a final forecast when judgment or automation changes the number;
7. connect the metric to the decision cost rather than to a generic accuracy target.

This order makes a tidy average harder to manufacture accidentally. It also makes disagreement more useful. Two people can disagree about the right decision while still agreeing on which rows, definitions, and errors the decision rests on.

Three conclusions to resist

First, do not infer intent from bias. A positive average may result from a changed mix, a new segment, a different issue horizon, or missing low outcomes. The number flags a pattern. It does not identify the motivation.

Second, do not infer operational safety from low mean absolute error. A small average can hide a tail that matters to a capacity constraint or a major account. Tail analysis is not optional when the downside is concentrated.

Third, do not infer decision quality from metric improvement alone. A more accurate forecast may arrive too late, use a definition that does not match the decision, or improve a measure while worsening a service, cash, or resource outcome. The metric needs a declared use.

How to diagnose a decline in sales forecasting before it becomes a growth problem

A sales-forecasting decline is easy to misread as a model problem when the object, horizon, or denominator moved first. Improvement starts by naming what got worse and for which decision, not by retuning accuracy alone. Use the following checks before changing the model, widening the pipeline, or opening a new forecasting workstream:

1. Name the declined object. If MAPE, bias, coverage, or win rate moved, say which one and for which horizon and decision. Do not treat a mixed dashboard drop as one forecast failure.
2. Separate accuracy from bias and mix. If aggregate accuracy looks stable while stage or segment errors offset, improve the row-level audit before changing the model. Do not celebrate a mean that hides canceling misses.
3. Check the denominator. If the forecast base, open pipeline, or eligible population changed, restate the base before blaming the algorithm. A moving denominator is not an improvement signal.
4. Match the fix to the failure. A definition gap needs a clearer forecast object; offsetting stage errors need a row-level audit; a dashboard that cannot say what changed needs diagnosis before a new model. Do not open an improvement program that answers a different question.
5. Refuse a universal improvement checklist. These checks stay inside the pair’s evidence limits. They do not provide a private operating diagnosis, accuracy benchmark, or portable uplift from a new model.

How should sales forecasting reconcile with finance and the P&L?

Operating forecast dashboards and the P&L answer different questions. A coverage multiple or MAPE can move while booked revenue does not, and a finance close can look healthy while the forecast object was never the cash line finance will use. Reconcile the two before changing targets, hiring plans, or model ownership:

1. Name the forecast object and the P&L line separately. Record which dashboard field is being governed and which income-statement or cash object finance will use to evaluate it. Do not treat pipeline coverage or stage probability as booked contribution by default.
2. Keep the horizon and close calendar aligned. If the forecast horizon and the finance close cut-off differ, state both before reading a miss as a model failure. Timing mismatch is not an accuracy result.
3. Separate bias from cash risk. If optimistic bias or canceling stage errors hide risk that finance must fund, keep the row-level audit visible beside the P&L bridge. Do not clear a finance review with an aggregate accuracy score alone.
4. Match the decision to both sides. A hiring or capacity decision needs a named forecast object; a margin or cash decision needs the finance line. Do not open one workstream that answers the other team’s question.
5. Refuse a private operating diagnosis as a universal bridge. These checks stay inside the pair’s evidence limits. They do not provide a private P&L result, accuracy benchmark, or portable uplift from a new model.

How should teams calculate, govern, and checklist sales forecasting?

Sales forecasting is a decision system, not a single accuracy number. Name the unit, horizon, and finance use before calculating or interpreting a dashboard. Compare sales forecasting to revenue forecasting when the question is pipeline conversion versus recognized revenue. Failures cluster when offsetting errors cancel into a flattering score. A checklist is a diagnostic prompt.

Boundary

The claim here is narrow: forecast accuracy cannot be interpreted responsibly until signed error, absolute error, distribution, horizon, denominator, and decision cost remain visible. The two cited studies provide bounded demand-planning and supply-chain evidence for separating accuracy, bias, and adjustment direction. They do not validate a universal B2B sales-forecast rule. A commercial transfer hypothesis would need its own baseline, event definition, repeated outcomes, and test design.

References

1. Fildes, R., Goodwin, P., & De Baets, S. (2025). Forecast value added in demand planning. International Journal of Forecasting, 41, 649–669. https://doi.org/10.1016/j.ijforecast.2024.07.006
2. Fildes, R., Goodwin, P., Lawrence, M., & Nikolopoulos, K. (2009). Effective forecasting and judgmental adjustments: An empirical evaluation and strategies for improvement in supply-chain planning. International Journal of Forecasting, 25(1), 3–23. https://doi.org/10.1016/j.ijforecast.2008.11.010