0

I am trying to forecast a very typical sales data. I have tried Arima, ETS, Holtwinters and even neural networks but I can't get a model with more than 40% accuracy [Absolute sum of(forecast-Actual)/Sum of Actual]. Can anyone suggest me how to proceed?

Here is the data https://drive.google.com/open?id=1uugpk53hMyF8ncl30vkkwhuJeVYFEUql

Nish7416
  • 1
  • You've got a number of variables listed. Are you trying to predict all of them? – TheEnvironmentalist Oct 01 '19 at 05:04
  • 1
    Just from eyeballing your data, you seem to have many intermittent series. These are typically very hard to forecast more accurately. If all you care about is minimizing the wMAPE, consider forecasting a flat zero for all highly intermittent series ([Kolassa, *IJF*, 2019](https://doi.org/10.1016/j.ijforecast.2019.02.017)). – Stephan Kolassa Oct 01 '19 at 06:23
  • 1
    @TheEnvironmentalist: I would presume each line to be one stock-keeping unit (SKU), and each column to be a month, so yes, each line would need to be forecasted separately. This is a very common setup in sales forecasting. There may be hierarchies involved that might make forecasting easier, but we don't know. – Stephan Kolassa Oct 01 '19 at 06:25
  • Yes, each line is one stock-keeping unit (SKU) and each column is a month. – Nish7416 Oct 01 '19 at 11:36
  • @StephanKolassa Sure I will try that. Thanks. Can you give me code in R programming for forecasting this data? – Nish7416 Oct 01 '19 at 11:45

0 Answers0