Update and Forecast Model - ValueError: p < 0, p > 1 or p contains NaNs about pybats HOT 3 CLOSED

Hi Nikita,

I apologize for the late reply, I didn't see this issue until today. I'd be happy to take a look at your data and work through the issue with you, if you'd like to send it to me at [email protected].

Thanks,
Isaac

from pybats.

from pybats.

Hi Nikita,

Thanks for your response, I'm glad that you've been using the package and that it's working well for you! Item-interaction effects are an interesting question, it's something that I thought about a little bit, but never had the opportunity to work on much.

If you'd like to work with something that already exists within PyBATS, I'd recommend using latent factors to capture interactions among items. The latent factor is a regression predictor that is a random variable. To give an example of using a latent factor, this paper uses a model on total sales to create a latent factor which is a forecast of day-of-week and holiday effects. If the same latent factor is used in multiple item models, then there will be dependence in the forecast. You can model the joint forecast distribution across items either with a Copula model using this function, or by simulating the latent factor, as is done in this paper.

To focus on item-interactions in particular, here is one simple idea: Simulate from the forecast distribution of one item, and then use those simulated values as predictors for another item. Pybats has functions to work with these simulated values, in this module.

In terms of developing PyBATS further, I would love to have you and others contribute to the package! I am no longer in graduate school, so I don't have as much time to actively develop, but I can certainly help out if there are things you would like to add. Please let me know if you have ideas for additional models, and feel free to submit pull requests.

Best,
Isaac

from pybats.