Multivariate Time Series


I’ve been thinking how to go about a specific problem and how I believe can be solved using bidirectional RNNs.

Let’s say I want to predict the sales for a given time step t. For the sake of argument we define each time step to be a day.

Suppose that I have the actual sales per day which can be used to train the network. Additionally, I have another variable which tells us if a given day is “special”. It could be anything, you can think of it like Christmas, New year’s etc. This variable could be binary [0,0,0…,1,0,0,1,1,0,…] or even continuous if you put different weight on different days.

Now let’s assume that the sales on a given day t are affected by the previous days sales t-delta and if a day is special or not.
Also assume that future special days t+delta’ also affects the sales on day t think of it like people buying in advance to prepare the special day ).

My main problem is how to go about the specifics of training the network and making predictions. You would want to use the variable “sales” up until time step t ( because that is what you would have in a real scenario ) but use the other variable up until time step t+delta since that is information that you already know.

How to go about this ? Pad the variable sales with zeros up until t+delta ? Mask the variable somehow ?

Thank you. I hope I stated the problem clearly enough.


Please go through this notebook and this course for starters.


Can you explain what do you mean by t+delta and t-delta?

If you are using a time series algorithm, you could use GRU which can process sequential data such as text, speech, and time-series data in your case.

where in you can create a model Γu is your updated Gate and you can assign the model Γu=0 in case you want your model algorithm to forget the day as it is not a special day like Christmas and use Γr=1 in case you want the referral gate to be assigned as per your special day for sales.

Read or listen videos on GRU it will help you address time series data concern.


Can I know about this data you are going to use or have used to make predictions.
This part window Lp is defined on what criteria.

I always have a concern/question regarding this special days, like Christmas Day is a fixed date 25th Dec but special day or festival days especially in India are based on cycles of Sun and moon movements like this year Diwali might be on October 20 but next year it can be on 2 Nov. How are you planning to address this issue if you are planning a global sales analysis.

If you are asking future perspective of making predictions you could create an algorithm based on how these special days are decided based on that days planetary positions, astral relation between each moon cycles.

Also can I know what are these other variables from future you are looking to predict.

One needs to understand in sales, the advantages and disadvantages of analysis

like the plus point is each cycles can be divided into a year which can again be divided into months. Further the days can be classified into fixed special days like New year, Labor Day, Christmas days etc!!! and other special days which will be effected with your times/days in relation to the moon cycles (which is where you need to find a relation between these special days and moon cycles like Diwali is celebrated on amavasya which is no moon day or new moon day. Same wise On full moon : Holi, budhha purnima, guru purnima, Rakshabandhan, Guru Nanak Birthday. Further actual days is will be your usual days or casual days.

I hope you understood my idea.