Using the R-MAPE index as a resistant measure of forecast accuracy

Area of Science:

• Statistics
• Econometrics
• Machine Learning

Background:

• Mean Absolute Percentage Error (MAPE) is a common metric for evaluating forecast accuracy.
• MAPE's validity is compromised by right-skewed error distributions and outliers, leading to overestimation.
• Existing methods lack robustness in the presence of skewed data and outliers.

Purpose of the Study:

• To introduce a robust alternative to MAPE for assessing forecast accuracy.
• To address the limitations of MAPE in handling skewed error distributions and outliers.
• To propose the Resistant MAPE (R-MAPE) as a more valid goodness-of-fit measure.

Main Methods:

• Utilized Artificial Neural Network (ANN) and Autoregressive Integrated Moving Average (ARIMA) models for time series forecasting.
• Developed the Resistant MAPE (R-MAPE) using the Huber M-estimator.
• Compared the performance of R-MAPE against MAPE on datasets with skewed error distributions and outliers.

Main Results:

• MAPE was found to overstate the population parameter in the presence of statistically significant right skew and outliers.
• R-MAPE demonstrated a more accurate reflection of the population parameter compared to MAPE under these conditions.
• The Huber M-estimator effectively mitigated the overstatement issue caused by skewed data.

Conclusions:

• R-MAPE is a suitable and valid alternative measure for forecast accuracy.
• R-MAPE provides a more reliable assessment of forecast performance than MAPE, especially with skewed error distributions.
• The proposed R-MAPE enhances the reliability of forecast evaluation in statistical modeling.