# Add External Predicted Feature Time Series

Hi,

I was wondering if anyone can help. I understand how to predict future values based on history. But how do I include externally predicted future features. So what will be the profit in week 5 based on historical data and the price of a widget being x in the forecast week 5.

Any help would be really appreciated.

Best regards,
Seamus

Externally predicted mean you want the model to use features not in the lists ? Or you want multiple model to work together ?

Hi,

Thanks for looking at this. So I work for a co-op and we set the milk price we pay to the farmers. In the calculation of lets say profit for week X, we would want to predict the profit in that future week based on the milk price we set. So we can see the impact that milk price will have.

Should I be making feature predictions for that week. Then add those features and the milk price to predict the profit lets say?

Best regards,
Seamus

So u set the price so u know the profit when u ave the milk quantity of next week (my family own lands and they produce also milk)

week 1 price 0.20\$ milk quantity x (this is the real factor to predict)
week 2 price y milk quantity x week1

is not a model thing to predict a calculation it will be easy top predict a scalar based on the past production volume. Maybe is very easy just with a LSTM with TimeStep 7 (one each day).

So looking the past model can predict the Volume in of milk and this way u can compute the price and the profit in advance. But since u set the price and i guess u also have a CAP on the volume imposed by the state u live in all can be simpler.

Hi Igor,

Thanks for coming back. Yeah that is sort of it. But the milk price has a big impact. We have contracts with customers and their product costs are set for a number of week. And we have inventory. And then we have differences in supply depending on the time of year. So changing the milk price at different times of year will have a different impact. There can be quite a lot of features that will have an impact.

So if I wanted to predict money coming in, in week x based on a milk price of y, There will be a lag in the impact and calculus and machine learning should be able to predict this lag until all contracts have been adjusted to represent the new milk price paid.

Is there a way for week x to let it predict the receipts based on historical data and the externally entered milk price for week x, since we will be setting it?

Best regards,
Seamus

We should get the dataset and see what feature create to optimize it. Is a regression problem with constraints so need feature engineering and a bit of study.

I think LSTM maybe with attention mechanism can be. A good base. If you want to plan more than a week you can even set steps to 28. Feel free to reach out in DM is u need more help since data is sensitive matter.

This (real life) problem looks interesting, although I’m not sure I fully understand your problem setup/data. It would be nice if you could comment more about it.

Hi Igor,

That would work. Thank you. I am from dev background and new to machine learning. If I put some sort of decay of relevance on the feature that would work. So older properties have less impact.

Thanks,
Seamus

Did u made to solve the issue or are u still at same point ?

Hi Igor,

Sorry, only seeing this now. I took your advice and have just been studying up until this point.

There are a few potential fixes I can use. Like a wide Keras with 2 side by side models and a concatenation. A RNN where I just add known value into the input when creating the next prediction in the window. A bi directional RNN. Still just studying but I should be able to get a solution.