Attached is a Table summarizing the MAPE and stationary R-squared values obtained for alternative exponential smoothing forecasting models for different variables. In general, the model with the lowest MAPE is not corresponding to the model with the highest stationary R-squared value. What are the possible reasons behind this discrepancy? Is it a must to have the lowest MAPE value for the model with the highest stationary R-squared value?
> Is it a must to have the lowest MAPE value for the model with the highest stationary R-squared value?
Not necessarily. In fact the MAPE measure has many disadvantages and can be misleading. In this chapter you will find an overview of the potential problems when using MAPE:
Chapter Forecast Error Measures: Critical Review and Practical Recommendations
R2 is an in-sample metric. If used to compare models, it can lead to overfitting.
In the chapter above we recommend to use the MAE metric for a single series or the AvgRelMAE metric if you have many time series. The MAE/AvgRelMAE metric will correspond to the best model if you use the posterior median as your point forecast.
This is very common. In this case you have to select the model with the lowest MAPE value. Or even you can use MAE.
- Badges
- Science topic
- Similar topics
- Philosophy
- Philosophy Of Science
- Reasoning