Suppose I am supplying $N = 1000$ vendors, and I am looking for a way to predict their demand for my product over $T = 90$ days. Concretely, I hope to take some features for each vendor, such as their size, geographical location etc., as well as their historical demand over some window of $w$ days, and predict their demand on day $w + 1$.

As input data for this problem, I have a dataset on vendor demand over a 90 day window, but the crucial issue is that for any particular vendor, they may arbitrarily report late. For example, if I queried vendor A on day 10 for their demand, they may only report their demand 2 days later, on day 12 (and so I am missing data on day 10 as well as day 11). This behavior varies across vendors, randomly, so that some vendors might report 1 day late; others may report 6 days late.

My question is: Can any predictive model trained on such a data set be meaningful, given this missing data mechanism? Assume that at a population level, I have some amount of demand data for any given window $w$ (but that this is not necessarily true for at the level of particular vendors). Moreover, if I wanted to avoid imputation, as doing this with $N = 1000$ seems to potentially introduce a lot of error, what would be a principled way to preprocess this dataset for the prediction task?

There are similar questions here and also here, but do not have the same type of missing data mechanism.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.