Total DL beginner here so forgive my stupidity, but I feel like I missed something – Once the model is trained, validated and tested, but how do you get the predicted ‘sales’ for the next day from the rossman model?
Same issue here. I’m not sure if the model is even ready for it. It seems like the other examples can produce predictions on new data, but not in the case of ColumnarModelData. I could be wrong but I’ve been looking for days with no results.
Admittedly, I am a little disappointed that no in the forum has responded. I have been researching this a fair amount in the last month and have come up with some direction (keep in mind I am new at this…). Check out this article How to Convert a Time Series to a Supervised Learning Problem in Python.
The summary of what he is saying:
- you have a series of data points: for instance, open, high, low, close (OHLC) for a stock price.
- Given today’s OHLC, the goal is to predict tomorrow’s close
- shift the close value back so that you have O(t), H(t), L(t), C(t), C(t+1), where
tis time step
- feed OHLC at time
Xand C at time
In the article the author provides a method
series_to_supervised which does the sifting for you, it can add multiple time steps to each row of data points for prediction, as well as multiple lag(past) data points to each row for additional data (although the model seems to over-fit if you add to many lag points).
From what I understand this is a many to one approach. Sequence models can also be many to many (similar to what is use in the language prediction models) which is what I am trying to figure out i.e. for the Rossman model, given sales data for a year, what would be the next month prediction of sales. I have not put all the pieces together in my brain yet as to how to make it work.