I’m wondering if I use Backward Looking Transformations, what should I do about the missing values? Is it better to drop the column with N/A or it’s better to fill it with zero or mean?
It is generally better to fill it with the mean, if you are using lightgbm models, you can even leave the N/A as it can deal with it automatically.
How should we deal with NaN values when working with deep learning models? We cannot fill it with the mean because of data leakage issues.
You should be fine by filling in using the mean in the training data. Another thing you can do is forwardfilling df.ffill() which only use past information for filling.