Course contentsShow
← Machine Learning and Deep Learning
Lesson 473 of 3,53811. Model Evaluation MetricsFree lesson

Mean Absolute Percentage Error (MAPE)

Understand MAPE for scale-independent evaluation, its percentage interpretation, and failure modes with zero or near-zero values.

Mean Absolute Percentage Error (MAPE)

What you'll learn: You'll understand MAPE as a scale-independent regression metric that expresses prediction error as a percentage, along with when it works—and when it dangerously fails.

The Core Idea

MAPE measures how far your predictions are from actual values, but expresses the error as a percentage rather than absolute units. This makes it incredibly intuitive: "Your model is off by 15% on average" is much easier to interpret than "your MAE is 3,472 units."

The formula calculates the absolute percentage error for each prediction, then averages them:

MAPE = (1/n) × Σ |actual - predicted| / |actual| × 100

Think of it like measuring how wrong your GPS estimated arrival time is. If you arrive in 100 minutes instead of the predicted 90, that's a 10% error—regardless of whether the trip was 90 minutes or 9 hours.

Why It's Useful

Scale-independent evaluation means you can compare models across different datasets or target ranges. A MAPE of 5% tells you the same story whether you're predicting housing prices (100K–100K–1M) or daily sales (10–100 units).

Critical Failure Modes

MAPE has a devastating weakness: it breaks down when actual values are zero or near-zero. Division by zero is undefined, and tiny denominators create massive percentage errors. If your actual value is 1 and you predict 5, that's a 400% error—but if actual is 1000 and you predict 1004, it's only 0.4%.

This creates asymmetric penalties: MAPE punishes over-predictions less than under-predictions, making it biased toward models that predict lower values.

Key Takeaway: MAPE provides intuitive percentage-based error measurement that's scale-independent, but fails catastrophically with zero or near-zero actual values and introduces bias toward under-prediction.