Right Windowing for time series forecast of sales
Hello Community,
I got 5 years (2014-2018) of monthly sales data on a specific product. I would like to build 12 Random forest models to predict each month of the year 2018.
1. Model to predict January 2018
2. Model to predict Feburay 2018
...
12. Model to predict December 2018.
So each model will be trained to predict one specific month. I have tryed a lot and made some predictions on the training dataset and evaluated them on the test dataset, but I am still unsure about the windowing. Should it look like this to train my models right?
|
Last consecutive Month in window |
Sales + 1 horizon (label) |
Sales-47 |
Sales-46 |
… |
Sales-0 |
|
48 |
y49 |
Jan 2014 |
Feb 2014 |
… |
Dec 2018 |
|
50 |
y50 |
Feb 2014 |
Mar 2014 |
… |
Jan 2018 |
|
… |
… |
… |
… |
… |
… |
|
60 |
y59 |
Dez 2014 |
Jan 2015 |
… |
Nov 2018 |
Thank you for any suggestion.
Find more posts tagged with
Sort by:
1 - 2 of
21
Altair Employee@Daniel_Tu is there a reason for doing it with Random Forest and 12 Models? It seems you could use time series analysis and Holt Winters oeprator with it your could predict the next 12 months based on the data you already have of the 4 previous years.
@MarcoBarradas I already used some other methods to predict my next year. Holt Winters, ARIMA, naive Forecasting and so on. I would like to compare these methods, that ist the reason for asking this. Setting my forecast horizon using Random forest is not as easy as it is with the times series methods.
Thank you for your response.