Version 1.8 of tsintermittent has been submitted to CRAN and should be shortly available for download. Amongst various new checks on inputs to better accommodate handling multiple time series with data frames, a new option has been added to
method="auto" two things will happen:
idclass(...,type="PKa")will be called to classify the time series and select for each one the appropriate forecasting method between Croston’s method, SBA and single exponential smoothing (for details on the classification see the documentation of
- Each time series will be forecasted using the selected forecasting method. Any parameters are optimised per time series.
Some things to keep in mind. The function
data.frc can accept additional inputs that are passed to the forecasting method used. The function is smart enough to distribute options that are only available to
sexsm appropriately. Also, you will get the same results if you use:
in which case
imapa is restricted to using only the original temporal aggregation level. However calling
method="imapa" instead of
method="auto" is substantially slower
, so the latter is recommended when handling multiple time series and you do not need to take advantage of temporal aggregation.
This paper empirically demonstrates that using a similar classification to select the best method for intermittent demand results in good forecasting performance. Although the good performance of the classification scheme was verified again in this paper, we also found that
imapa gave the most accurate forecasts. Nonetheless, the new option should allow to quickly implement either approach. My personal view is that the method selection issue for intermittent demand time series is far from resolved, as I demonstrate in this paper, but good progress is being done and should be used in practice.