When forecasting for group of products, is there a way to let statsforecast (and other Nixtla libraries) know that some of the products have incomplete sales history? Example, product 1 has full history, product 2 has NaN, NaN, …, Nan then sales history starts. I would imagine libraries won’t accept NaNs and filling them with zeros would also be incorrect as these product might have had sales history but due to forecasting / ERP system change sales history for some products was lost.

Hi @Valeriy currently StatsForecast doesn't have that functionality, so you'll need to use domain knowledge to fill the missing values as best as possible. What StatsForecast does allow is different start dates, so if product 2 has missing values, you can just remove them until the first non-zero value.

@Mariana Menchero thank you. ‘What StatsForecast does allow is different start dates’ is there example of how this can be done in one dataframe?

@Valeriy https://nixtlaverse.nixtla.io/utilsforecast/preprocessing.html

@Guillaume GALIE that’s a different thing, I am interested in how one can set up different start dates for statsforecast.

@Valeriy 🤔 I am not nixtla expert but isn't the default behavior for statsforecast with an individual model for each time serie ?

@Guillaume GALIE You might be right, it might be as simple as having different start dates for different products. I will check.

You can use the

fill_gaps

function to find out which are the missing values ​​that are not in the training data or in the testing data, for example if you have a monthly dataset 2020-01-01, 2020-02-01 2020-04-01 In this case, the model can send you an error that you have NaN, because you need month 2020-03-01.