Time series model with missing week end

[Abhishek Saha]

Hi,

In our time series problem we are having the dataset that contains the data only for week days.

So, in our case, if we are presently at on Wednesday.....

1 day future value means Thursday.

2 day future value means Friday.

3 day future value means Monday (not saturday)

4 day future value means Tuesday (Not Sunday).

But while working with time series problem, DataRobot can't exclude the week ends. So, if we provide the forecast window 1-40 days. Then practically we receive the forecasting of next 28 week days (because remaining 2X6=12 are weekends). But we need the forecasting for next 40 week days.

What should be the solution? Thanks for the help for your advice.

Thanks and Regards

Abhishek Saha

[abhishek.saha@molgroup.com]

@JessLin Thanks for your answer.

But I have tried both way (Day as well as Row). In both the cases the forecasting window number of days includes Saturday/Sunday. So, for 40 days forecasting window shows 28 days value.

[jess_lin]

Just to clarify - do you include weekends (have values for all 7 days of the week), or does your data exclude weekends (have values for weekdays only)? If the former, I would just set up a forecast window to predict 1 day out:

[zsfeinstein@yahoo.com]

Which of the two solutions presented here would you recommend for a scenario where we would like to predict just one day in advance? I also have a model without weekends and without major holidays.

Is there a set of instructions in how to properly set this up?

[JessLin]

Hi Abhishek,

There are a couple different ways you can handle this!

1) Switch time series to "row-based" mode, and set the forecast window to [1,40], and make sure that your prediction dataset only contains weekdays

(screenshot of how to set up the time series settings)

Pros: The "forecast distance" matches up to future # weekdays (eg "1 row ahead" will always equate to "1 weekday ahead" regardless of weekends)

Cons: you lose any date-related feature engineering (eg "week aggregation")

2) Stay in "day-based" mode, but increase the forecast window to [1,56]

(screenshot of what the output would look like - notice the missing weekends)

Pros: Keep date-related feature engineering, and you'll still forecast the next 40 week days

Cons: Your forecast distance doesn't precisely match up to future # of weekdays (depending on your forecast point, "forecast distance = 3" may mean 3 days ahead, or it may mean the next weekday)

Hope this helps!

- Jess