How is forecast accuracy measured?

Two figures, never one. Mean error captures bias — a persistent direction. Mean absolute error captures dispersion regardless of direction. Reporting only one is the most common measurement error in forecast review, and they call for opposite responses.

What is a good forecast accuracy?

Improving against your own history at a fixed submission point. Cross-company comparisons are meaningless because the submission timing, the categories and the definitions all differ.

Why is our forecast always wrong in the same direction?

Persistent directional error is bias, and it is correctable — apply the historical adjustment rather than asking for more accurate submissions. Persistent bias usually reflects how the forecast is used rather than what forecasters believe.

Does more frequent forecast review improve accuracy?

It addresses judgement error at the deal level only. Where the error is structural — stage probabilities that no longer match observed conversion — or incentive-driven, more review changes nothing.

Forecast Accuracy

How closely forecasts match actual results, measured consistently over time. The metric that separates disciplined forecasting from lucky forecasting.

How to measure it

Two figures, computed separately and reported together.

Mean error — actual minus forecast, signed and averaged. This is bias: a persistent lean in one direction.
Mean absolute error — the same differences with the sign discarded. This is dispersion: how far off a typical period is, irrespective of direction.

They demand opposite responses. Bias is systematically correctable by adjustment. Dispersion is not, and is reduced only by better information or better process.

Why one number is not enough

A forecaster who is 20% high one period and 20% low the next has near-zero bias and severe dispersion. One who is 10% high every period has meaningful bias and low dispersion. Reported as a single accuracy figure they can look identical, and the correct action for each is different.

Fixing the submission point

Accuracy is meaningless without stating when the forecast was made. A forecast submitted on the last day of a quarter should be near-perfect and says nothing about forecasting capability. Measure at a fixed point — the start of the period, and again at the midpoint — and compare like with like over time.

Locating the error

Test the structural layer first: for each stage, compare assigned probability against observed historical conversion. Divergence here is systematic and invisible at the deal level.
Test the judgement layer: compare earliest submitted forecast to outcome, grouped by forecaster. Consistent direction by individual is correctable with a personal adjustment.
Test the incentive layer: plot when forecast category changes happen. Systematic movement in the final week reflects how the forecast is used, not information about deals.
Attribute the total across the three. The largest source is frequently not the one currently receiving attention.

Where it misleads

Accuracy improves mechanically as the period progresses, so an unstated submission point makes any figure unfalsifiable.
Aggregate accuracy can be good while every constituent forecast is poor, when errors offset. Check the components.
Where the forecast also sets quota or evaluates performance, it is not purely a prediction, and its error partly measures that use.

RELATED TERMS

Pipeline Coverage

The ratio of open pipeline value to the target for a period. Answers whether there is enough in play to hit the number at historical conversion rates.

Win Rate

The proportion of opportunities that close won, measured against a defined denominator — which is where most of the disagreement lives.

Weighted Pipeline

Pipeline value multiplied by a probability assigned to each stage, producing an expected value rather than a gross total.

COMMON QUESTIONS

How is forecast accuracy measured?: Two figures, never one. Mean error captures bias — a persistent direction. Mean absolute error captures dispersion regardless of direction. Reporting only one is the most common measurement error in forecast review, and they call for opposite responses.
What is a good forecast accuracy?: Improving against your own history at a fixed submission point. Cross-company comparisons are meaningless because the submission timing, the categories and the definitions all differ.
Why is our forecast always wrong in the same direction?: Persistent directional error is bias, and it is correctable — apply the historical adjustment rather than asking for more accurate submissions. Persistent bias usually reflects how the forecast is used rather than what forecasters believe.
Does more frequent forecast review improve accuracy?: It addresses judgement error at the deal level only. Where the error is structural — stage probabilities that no longer match observed conversion — or incentive-driven, more review changes nothing.