Skip to content
QuorlyticTalk to a Data ExpertContact
Menu

Implementation guide

Accuracy tells you how far off a forecast is. Bias tells you which way.

A demand plan can report 92% accuracy for a quarter and still leave a warehouse full of stock nobody ordered. Both facts fit together once direction is measured: the forecast was close in size and wrong in the same direction every week. This guide sets out the few measures that catch that, how to compute them, and how to report them so a planner can act.

Direct answer

Forecast accuracy is usually reported as 100% minus WAPE: total absolute error divided by total actual demand, subtracted from 100%. Forecast bias is the signed total of forecast minus actual, divided by total actual. Accuracy says how far off a forecast is; bias says in which direction. A tracking signal flags persistent bias early, and forecast value added shows whether each step in the process beats a naive forecast.

Engineering reference guideUpdated Nov 10, 202613 min read

  1. 01Fix level and lag
  2. 02Compute accuracy
  3. 03Track bias
  4. 04Test each step against naive

The forecast accuracy formula

Forecast accuracy = 100% − WAPE, where WAPE = Σ|Forecast − Actual| ÷ ΣActual. With 1,695 units of demand and 73 units of total absolute error, WAPE is 4.31% and accuracy is 95.69%. Pooling the error before dividing is what keeps the figure defined when some periods have zero demand.

Some teams compute accuracy as 100% minus MAPE instead. That inherits every weakness of MAPE: it cannot handle zero actuals, it gives a slow seller the same weight as a best seller, and it rewards under-forecasting. The MAPE vs WAPE vs RMSE comparison sets out the arithmetic; for an accuracy figure people will be measured on, WAPE is the safer base.

Accuracy can go below zero. When total error exceeds total demand, as it can for intermittent items, 100% minus WAPE is negative. Report WAPE itself for those items, or floor accuracy at zero and say that you did, rather than letting a negative percentage appear in a review with no explanation.

Forecast accuracy and bias measures: how each is calculated, what it tells you, and what to watch for
DimensionFormulaWhat it tells youWatch for
Forecast accuracy100% − Σ|F − A| ÷ ΣAHow close the forecast was, weighted by volume.Goes negative when error exceeds demand.
WAPEΣ|F − A| ÷ ΣAThe size of the error as a share of demand.Hides a badly forecast low-volume tail.
Bias %Σ(F − A) ÷ ΣADirection: positive is over-forecast.Noise over one period; read it over several.
MADΣ|F − A| ÷ nTypical miss in units.Not comparable across products of different size.
Tracking signalΣ(A − F) ÷ MADWhether bias is building up over time.Sign depends on the error convention used.
Forecast value addedError of previous step − error of this stepWhether a step beats the one before it.Needs the forecast stored at every step.

Fix the level and the lag before computing anything

Measure accuracy at the level where the decision is made and at the lag the decision needs. A replenishment decision taken two weeks ahead for each product at each warehouse should be measured on the forecast as it stood two weeks before the week, at product-warehouse level. Any other level or lag answers a different question.

Level matters because errors cancel as you aggregate. The same forecasts can show a much larger error at product-store level than at national level, and neither number is wrong; they answer different questions. Reporting at a level nobody acts on is the most common way a planning team reports progress while shelves stay empty.

Lag matters because forecasts change as the date approaches. Comparing actuals with the latest forecast, made the day before, measures a forecast that was too late to change an order. Freeze a copy of the forecast at the decision lead time and measure that copy. Storing those snapshots is a data engineering task, and without it the accuracy figure flatters the process.

  • Level: the grain at which someone orders, schedules or allocates
  • Lag: the lead time between the forecast being used and the period it describes
  • Coverage: the share of items and volume included, reported beside every figure

Forecast bias: the direction of the error

Forecast bias is Σ(Forecast − Actual) ÷ ΣActual. A positive value means the forecast ran high and a negative value means it ran low. Unlike accuracy, bias keeps the sign of each error, so persistent over-forecasting, which builds excess stock, and persistent under-forecasting, which causes stockouts, become visible instead of averaging into ordinary error.

A single period's bias is mostly noise. What matters is whether the sign stays the same. A bias that alternates between positive and negative is random error; a bias with the same sign for six or eight periods is a process problem with a cause that can be found.

Common causes are specific and fixable: an uplift added for a promotion that was later cancelled, a discontinued product still being forecast, a calendar shift such as a moved holiday, demand history that recorded constrained sales during stockouts as low demand, or a planner adjustment applied in the same direction every cycle. Accuracy metrics show all of these as generic error and give no clue where to look; bias points at them.

The tracking signal: catching bias early

The tracking signal is the running sum of errors divided by the mean absolute deviation: Σ(Actual − Forecast) ÷ MAD. While errors cancel out it stays near zero. When they keep the same sign it grows by about one unit per period. A common rule of thumb investigates the forecast when the signal moves beyond ±4.

The sign convention matters and varies between tools. With error defined as actual minus forecast, a negative tracking signal means the forecast has been running high; some planning systems define error the other way round and flip the sign. State the convention next to the chart.

The ±4 limit is a convention rather than a law. Tighter limits catch bias sooner and raise more false alarms; looser ones stay quiet longer. What matters is that the limit is fixed in advance and that crossing it triggers a review with an owner, not just a red cell in a report.

Forecast value added: does each step beat doing nothing?

Forecast value added (FVA) compares the error of each step in the forecasting process with the error of the step before it, starting from a naive forecast such as last period's actual. A positive FVA means the step reduced error; a negative FVA means the process would have been more accurate without it.

The usual steps are a naive forecast, a statistical model, a planner adjustment and a consensus or sales override. Measuring FVA requires storing the forecast after every step, at the same level and lag, which most teams do not do until they decide to measure it.

FVA turns a debate about whose forecast is better into a measurement. It is also the fairest way to judge a new model: a sophisticated model that cannot beat the naive forecast on your own data has not earned its operating cost, and a planner adjustment that consistently removes value is a training or incentive question rather than a modelling one.

Worked example · hypothetical organisation

Worked example: one product, eight weeks, three forecasts

Over eight weeks with 1,695 units of demand, a naive forecast has a WAPE of 7.67%, a statistical model 4.31% and the consensus forecast after a planner override 8.26%. The model adds 3.36 points of accuracy over naive; the override then removes 3.95 points and adds a +8.26% bias, which the tracking signal flags in week 4.

The naive forecast misses by 130 units in total, so its WAPE is 130 ÷ 1,695 = 7.67%, with a small negative bias of −1.77%. The statistical model misses by 73 units: WAPE 4.31%, accuracy 95.69%, and a bias of +0.29%, which is effectively none. Its tracking signal moves between −1.27 and +0.22 across the eight weeks and never approaches the limit.

The consensus forecast adds between 15 and 20 units to every week. Its total error is 140 units, a WAPE of 8.26%, and because every error has the same sign its bias is also +8.26%. The tracking signal falls by exactly one unit a week, −1, −2, −3, and reaches −4 in week 4, half way through the period. A team watching only accuracy would see 91.74% and move on.

The forecast value added reading is the useful one: the statistical model is worth 3.36 points over naive, and the override costs 3.95 points against the model, leaving the final forecast slightly worse than repeating last week's sales. The fix is not a better model. It is a conversation about why every adjustment goes up.

Weekly actual demand for one product, with the naive, statistical model and consensus forecasts for each week
DimensionActualNaive (last week)Statistical modelConsensus after override
Week 1200195208215
Week 2210200200225
Week 3190210201210
Week 4220190212235
Week 5205220214225
Week 6230205221245
Week 7215230226235
Week 8225215218245
Tracking signal · eight weeksStatistical modelConsensus after overrideAction limits ±4
+40-4-8
Week 4: limit reached
  1. W1
  2. W2
  3. W3
  4. W4
  5. W5
  6. W6
  7. W7
  8. W8
  1. NaiveWAPE 7.67%baseline
  2. Statistical modelWAPE 4.31%+3.36 points versus naive
  3. ConsensusWAPE 8.26%−3.95 points versus the model

Report a small, fixed pack at the decision level

A forecast accuracy report that people act on has five numbers, all at the decision level and lag: accuracy as 100% minus WAPE, bias, items breaching the tracking-signal limit, coverage, and forecast value added by step. Keep the definitions fixed for long enough that a change in the number means a change in the forecast.

The pack is deliberately short. Every additional metric gives a team one more number that can improve while the business does not, and a long list of measures tends to be read selectively. Five numbers, read the same way every cycle, show whether forecasting is getting better and where it is not.

  • Accuracy: 100% minus WAPE, by segment as well as in total
  • Bias: signed, with the sign convention printed on the report
  • Exceptions: the items whose tracking signal crossed the limit this period, with an owner each
  • Coverage: the share of items and volume included in the figures
  • Forecast value added: naive, model, adjustment and consensus, side by side

When accuracy is the wrong target

Some demand cannot be forecast precisely: intermittent spare parts, new products with no history, and demand driven by one-off events. For those, chasing a better accuracy figure wastes effort. Measure service level and stock instead, and manage the uncertainty with safety stock, flexible supply or a forecast range rather than a single number.

Accuracy is also only worth improving if it changes a decision. If order quantities are still set by a fixed rule, a more accurate forecast produces a better report and the same orders. Connecting the forecast to the demand forecasting and planning workflow that uses it is where the value sits, and turning a prediction into a decision covers that step.

Finally, none of these measures is meaningful if the actuals are wrong. Sales recorded during a stockout understate real demand, and a forecast measured against them will look biased high when it was right. Separating demand from constrained sales is a data quality question before it is a forecasting one.

Questions and trade-offs

Question

What is the formula for forecast accuracy?

The most robust version is forecast accuracy = 100% − WAPE, where WAPE = Σ|Forecast − Actual| ÷ ΣActual. It pools the error before dividing, so periods with zero demand do not break it and large items count in proportion to their volume. State the level and lag it was measured at, because both change the number.

Question

What is a good forecast accuracy percentage?

There is no universal figure. Accuracy depends on how volatile demand is, the level of aggregation and the forecast lag, so a figure that is strong for one product at one store can be weak for a national monthly total. The meaningful comparisons are against a naive forecast on the same data and against your own history at the same level and lag.

Question

How do you calculate forecast bias?

Forecast bias = Σ(Forecast − Actual) ÷ ΣActual, expressed as a percentage. A positive result means the forecast ran high and a negative result means it ran low. Read it over several periods: a bias that keeps the same sign for six or more periods points to a specific cause worth finding.

Question

What is a tracking signal in forecasting?

A tracking signal is the running sum of forecast errors divided by the mean absolute deviation. It stays near zero while errors cancel out and grows steadily when they keep the same sign, so it flags persistent bias early. A common rule of thumb reviews a forecast when the signal moves beyond ±4.

Question

What is forecast value added?

Forecast value added (FVA) measures whether each step in the forecasting process reduces error compared with the step before it, starting from a naive forecast. It needs the forecast stored after every step. A negative FVA means a step, often a manual adjustment, made the forecast less accurate than it would have been without it.

Question

Should forecast accuracy be based on MAPE or WAPE?

WAPE, in most cases. Accuracy based on MAPE cannot handle zero-demand periods, gives a slow seller the same weight as a best seller, and penalises over-forecasting more than under-forecasting. WAPE is still a percentage that reads the same way to a non-technical audience, without those distortions.

Technical references

Share

Your next data decision

Start with a clear problem.
Build what comes next.

Bring the business question.
We'll help shape the engineering brief.