Does positive bias always mean overforecasting?

No. The interpretation depends on whether error is forecast minus actual or actual minus forecast.

Can an unbiased forecast be inaccurate?

Yes. Large positive and negative errors can cancel.

Will adding past mean error solve the problem?

It recenters that historical sample, but future improvement depends on stability and independent evaluation.

AI Glossary / Statistics and probability / Intermediate

Forecast bias

Definition

Forecast bias is a systematic tendency for predictions to be too high or too low relative to the quantity they are intended to forecast. A common diagnostic is mean signed forecast error. With error defined as actual minus forecast, positive mean error indicates underforecasting and negative mean error indicates overforecasting. Some systems reverse that sign convention, so the formula must accompany the label. A finite sample's average error estimates directional imbalance; a nonzero value alone does not prove a persistent population-level bias.

What it means

Bias concerns direction, while accuracy measures such as mean absolute error concern the size of misses. A forecast can have little average signed error yet make large mistakes because positive and negative errors cancel. Conversely, a modest but consistent directional error may reveal a correctable problem. The target also matters: a forecast intended to estimate a conditional mean has a different calibration objective from a median or high quantile used for an asymmetric decision. Calling every departure from zero mean error a defect ignores these differences in what the forecast is meant to represent.

How it works

Match each forecast to the actual outcome for the same target, horizon and information cutoff. Compute signed errors under a declared convention and summarize them over an appropriate evaluation period. Examine the sequence and relevant segments as well as the overall average. For conditional-mean forecasts, evidence of stable mean error can motivate an adjustment or model revision. Estimate any correction using information available before its evaluation, then test it on later outcomes. Uncertainty assessment should allow for serial dependence or overlapping horizons where present. Distinguish true out-of-sample forecast errors from fitted residuals, whose average may be constrained by the estimation method.

Three practical examples

Demand planning: investigate whether a model repeatedly predicts less demand than is later observed at the replenishment horizon.
Workload forecasting: compare directional errors by location to detect offsetting overprediction and underprediction hidden in an overall mean.
Model monitoring: check whether a previously calibrated forecasting process develops persistent errors after changes in products, customers or operations.

Think of it this way

Think of a weighing scale that repeatedly reads slightly low. That is a directional problem. A different scale might swing equally far above and below the true weight; its average error could be zero while individual readings remain unreliable. Forecasting needs both kinds of checks.

Worked example

Consider hypothetical actual outcomes [105, 112, 108, 115] and forecasts [100, 110, 100, 110]. Actual-minus-forecast errors are [5, 2, 8, 5], with mean five. The sample therefore shows average underforecasting of five units. Its mean squared error is (25 + 4 + 64 + 25)/4 = 29.5. Adding five to each forecast makes the same-sample errors [0, -3, 3, 0], with mean zero and mean squared error 4.5. The reduction of 25 equals the square of the original mean error. This is an arithmetic demonstration of recentering the observed errors, not evidence that the correction improves future forecasts: it was estimated on these same outcomes. Separately, errors [-20, 20] have zero mean but mean absolute error 20, illustrating why absence of net directional error does not imply accuracy.

Illustrative case study

Fictional case: Willow Planning Lab notices repeated underprediction for one product category. Analysts first verify the forecast timestamps and discover that a recent promotion variable was missing from the production data feed. They repair the input process and evaluate later forecasts, rather than automatically adding the recent average error forever. They retain a bias chart alongside absolute-error measures. This invented case demonstrates that a directional diagnostic can prompt investigation, while the cause and effectiveness of a remedy require additional evidence.

Common misconceptions and mistakes

Do not confuse forecast bias with social or demographic fairness bias; they are distinct concepts even when both deserve investigation. Do not interpret zero aggregate error as calibration for every subgroup or horizon. Avoid testing a correction on the same observations used to estimate it and calling that independent validation. A change in the actual measurement definition can mimic forecast bias. Percentage-error measures introduce additional denominator effects and may be undefined at zero actuals. An intentionally high service quantile should be judged against its quantile target, rather than automatically forced to have zero mean error.

Business use

Businesses can monitor directional error to identify planning problems that may accumulate into shortages, excess inventory or misplaced capacity. Keep forecast versions and measurement definitions stable enough for comparisons. Report the sign convention and units, and distinguish an observed recent average from a well-supported persistent pattern. Evaluate any correction at the horizon where decisions are made and inspect relevant segments. Choose the forecast target in relation to operational costs, then use evaluation measures consistent with that target. Bias monitoring is most useful when it leads to diagnosis rather than a mechanical adjustment disconnected from changing conditions.

Check your understanding

Under actual-minus-forecast errors, a mean error of negative six is observed. What direction does that indicate?

Answer: The forecasts exceeded actuals by six units on average in that sample. Persistence and uncertainty still need investigation.

Related terms

Calibration
Mean Absolute Error
Forecast horizon
Conditional means
Mean error
forecast accuracy
residual diagnostics
conditional mean
quantile forecast
out-of-sample evaluation

Learn next

Learn how a forecast's target determines its appropriate loss function. Then study rolling-origin evaluation and dependence in forecast errors before assessing bias corrections or comparing performance across horizons.

Loss function

Sources

Hyndman and collaborators, forecast package accuracy documentation: https://pkg.robjhyndman.com/forecast/reference/accuracy.default.html Hyndman and Athanasopoulos, Residual diagnostics: https://otexts.com/fpp2/residuals.html Gneiting, Making and Evaluating Point Forecasts: https://arxiv.org/abs/0912.0902

Previous: Forecast attainment Next: Forecast horizon