The following are my learnings from the Hyndman-MASE.

  • Scale dependent errors
    • Mean Absolute Error
    • Geometric Mean Absolute Error
    • Mean Square Error
  • Percentage Error
    • MAPE
    • sMAPE - to avoid the problem with MAPE where it puts more penalty on positive errors as compared to negative errors
  • Relative error : Divide each error by the error obtained by using some benchmark method
  • Scale-free method: Scale the error by mean absolute naive forecast error. Mean of all scaled errors
  • MASE is suitable even when the data exhibit a trend or seasonal pattern
  • Typical values for one-step forecasts are less than 1
  • Multi-Step MASE are often larger than one as it becomes difficult to forecast as horizon increases
  • MASE works for all types of forecasting scenarios