Can prediction bias be negative?

Yes. Under prediction-minus-observation, it indicates average underprediction.

Is zero prediction bias the same as full calibration?

No. Overall means can match while predictions are miscalibrated within score ranges or groups.

Should probabilities or binary decisions be averaged?

Use the quantity relevant to the diagnostic and state it explicitly. Probability calibration generally concerns predicted probabilities and observed labels.

AI Glossary / Governance, ethics and safety / Intermediate

Prediction bias

Definition

Prediction bias is the difference between the average of a model's predictions and the average of the corresponding observed labels in a dataset. With the convention prediction minus observation, a positive value indicates overprediction on average and a negative value indicates underprediction.

What it means

This is a specific numerical diagnostic. It is not the same as a model's intercept, the bias component in a bias-variance discussion or unfair treatment of people. A model can have zero prediction bias by this measure while still making poor individual predictions or producing harmful disparities.

For numerical forecasts, the comparison uses predictions and observations in the same units. For a probabilistic binary classifier, averaging predicted probabilities gives an estimated event rate that can be compared with the average of zero-or-one labels. The comparison should use the same cases and observation window.

Be clear about what is averaged. Thresholded class decisions answer a different question from predicted probabilities. A well-calibrated probability model need not classify the same fraction of cases as positive at every chosen threshold.

How it works

Collect matched predictions and sufficiently mature outcomes for the population being evaluated. Compute the mean prediction, compute the mean label and subtract the second from the first. Apply any justified weighting consistently and document exclusions or missing outcomes.

Investigate both the overall value and relevant slices, such as time periods, operating settings or score ranges. Positive and negative errors can cancel in the aggregate. A zero overall difference therefore does not establish good calibration across scores or good accuracy for individual cases.

A nonzero result can arise from sampling variation, changes in the population, inadequate features, fitting choices or errors in the data pipeline. It is a signal to investigate, not a diagnosis by itself. Avoid mechanically forcing current predictions to match an old training average when the actual event rate has changed. Evaluate uncertainty and the validity of the outcome labels before deciding what correction is appropriate.

Three practical examples

A demand-forecasting team checks whether weekly predicted totals systematically exceed observed totals, alongside measures of individual forecasting error.
A probability model's average predicted event rate is compared with the realized rate on the same completed evaluation cases.
A monitoring team calculates prediction bias separately by operating region to detect opposite errors hidden by the overall average.

Think of it this way

Think of a weighing scale that reads too high on some objects and too low on others. Its average error can be zero even though many individual readings are wrong.

The analogy has limits: a model's errors depend on inputs, data and learning, rather than only a physical calibration setting. It illustrates why an average difference is a useful check but an incomplete quality assessment.

Worked example

Zero average bias, substantial errors

An invented forecasting model predicts 20, 40, 60 and 80 units for four locations. The actual values are 40, 20, 80 and 60. Both sets average 50 units, so prediction bias is 50 − 50 = 0.

Yet every location's forecast is wrong by 20 units. The mean absolute error is therefore 20. Overprediction at two locations cancels underprediction at the other two in the signed average.

If stock must be delivered to each location separately, the correct total does not make those errors harmless. The team should examine error magnitude, location-specific patterns and the consequences of shortages and excess stock. Zero prediction bias establishes only that the averages match on these four cases.

Illustrative case study

Fictional case: Elm Planning notices that its model now predicts more bookings than the service observes. The first suggestion is to subtract a constant from every prediction.

A review discovers that recent booking outcomes have not all completed their reporting window. Some cases labeled as non-bookings are merely unfinished. The team aligns the prediction and observation periods before deciding whether a model adjustment is warranted, then examines residual patterns and uncertainty.

This imagined case demonstrates why measurement and data maturity need checking before correction. It does not claim that all prediction bias comes from delayed outcomes or that a constant adjustment is never useful.

Common misconceptions and mistakes

Zero prediction bias proves high accuracy. Large opposite errors can cancel.
Prediction bias directly measures ethical fairness. It is a mean-difference diagnostic; fairness requires context-specific analysis.
The predicted average must always equal the historical training average. The relevant population and outcome rate may have changed.

Business use

Prediction bias can be a simple monitoring signal when predictions have a clear numerical interpretation and reliable outcomes become available. Report its units, time window, population and uncertainty, alongside error magnitude and relevant calibration checks.

Use it to guide investigation rather than automatically changing decisions. Review data quality, population shifts and model assumptions. Where the results affect resource allocation, look beyond the total: offsetting errors can still create shortages, delays or unequal service. Document the sign convention so different dashboards do not appear to contradict one another.

Check your understanding

Two regions have prediction biases of +10 and −10 units with equal numbers of cases. The overall bias is zero. Is regional performance necessarily satisfactory?

Answer: No. Opposite regional errors cancel overall. Their magnitudes and operational consequences still require evaluation.

Related terms

Learn next

Read calibration and mean absolute error. Calculate both the signed mean error and the mean absolute error for a small forecast set, then explain why they answer different questions.

Sources

Google Machine Learning Glossary, prediction bias: https://developers.google.com/machine-learning/glossary#prediction-bias. Google Machine Learning Crash Course, Prediction bias: https://developers.google.com/machine-learning/crash-course/classification/prediction-bias. Relevant returned primary documentation defines the average comparison and identifies diagnostic limitations. Worked calculations, reporting-window scenario and operational interpretation are original teaching material.

Previous: Prediction Next: Prediction by partial matching