How to handle date in Random Forest prediction?

Question

I want to predict the income for small business from 1/1/2018 to 1/1/2020 for a dataset from historical data(all my variables are numeric except for date) the start date of my data is 1/1/2012. The income is updated each 3 months (so on 1/1/2018 I have x income and on 30/3/2015 I have new income) I'm not sure how to handle the date in this situation.

I did the following:

1- read the data

2- analysis the data

3- convert date column to type date 4- sort date

5- split the data (training data from 1/1/2012 to30/12/2017 and test data from 1/1/2018 to 1/1/2020)

6- convert date back to numeric

7- normalize all my columns except for the date columns

8- perform RF

9- predict using test dataset

In this scenario I didn't change the date format but I feel like this might not be the best way to handle date format.

Answer 1

Random Forest needs to handle dates as numeric data, that's why you can use the day, the weekday, the month, the trimester and the year as separated fields.

In addition to that, if your data has a cyclic behavior, you can align your data to it.

For instance, if your data has a cycle of 3 months, you can have the following features: day of the trimester (from 1 to 90), weekday, trimester and year. This would help to find interesting patterns.

Note: after the random forest processing, you may want to convert the day of the trimester to yyyy-mm-dd for a better clarity.