Learn→Time Series Forecasting→Model Evaluation (MAPE, RMSE, Forecast Bias)→3 of 6

Time Series Forecasting • Model Evaluation (MAPE, RMSE, Forecast Bias)Medium⏱️ ~2 min

What is Forecast Bias and Why Does It Matter?

Definition: Forecast bias measures systematic over- or under-forecasting: Bias = (1/n) × Σ(forecast - actual). Positive bias means consistently forecasting too high; negative bias means consistently forecasting too low. Unlike MAPE and RMSE which measure error magnitude, bias measures error direction.
Why Bias Matters
Unbiased forecasts have errors that cancel out over time—sometimes high, sometimes low, averaging to zero. Biased forecasts systematically miss in one direction, causing cumulative business impact. Persistent over-forecasting leads to excess inventory; persistent under-forecasting causes stockouts. Bias compounds while random errors average out.
Bias vs Variance Trade-off
A model can have low RMSE but high bias (consistently off by small amounts in same direction), or zero bias but high RMSE (large errors that happen to cancel). Optimal forecasts minimize both. If forced to choose, bias may matter more for inventory planning where systematic errors accumulate, while variance may matter more for pricing where errors do not compound.
Detection Method: Plot cumulative sum of errors over time. Upward slope indicates positive bias (over-forecasting); downward slope indicates negative bias. Flat line with oscillation indicates unbiased forecasts with random errors.
Sources of Bias
Structural model limitations: linear models underforecast exponential growth. Stale training data: model trained on pre-promotion data under-forecasts promotional periods. Incentive misalignment: sales teams over-forecast to secure inventory allocation. Identify and address root causes rather than just adjusting forecasts.
Bias Adjustment
If bias is consistent and predictable, adjust forecasts: corrected_forecast = forecast - measured_bias. But this is a band-aid. Investigate why the model is biased and fix the underlying issue. Persistent bias often indicates missing features, model misspecification, or data quality problems.

💡 Key Takeaways

✓Bias measures systematic directional error: positive = over-forecasting, negative = under-forecasting

✓Bias compounds while random errors average out—critical for inventory where systematic errors accumulate

✓Plot cumulative sum of errors: upward slope = positive bias, downward = negative, flat = unbiased

📌 Interview Tips

1A model can have low RMSE but high bias (consistently small errors in same direction)

2Common bias sources: structural model limitations, stale training data, incentive misalignment

← Back to Model Evaluation (MAPE, RMSE, Forecast Bias) Overview