shap How to use SHAP on Multivariate Time Series Forecast?

How to use SHAP on Multivariate Time Series Forecast?

Open lorenzofantunes opened this issue 4 years ago • 5 comments

Hello!

I'm trying to use SHAP in time series forecast, let's say I have a multivariete time series with various exogenous time series, and I want to predict a univariete time series.

I want to use it to explain the past and the future (montlhy forecast), for example I want to forecast sales of a product based on historical montly price, market share, marketing investiment, etc.

My impression is that I should create a wrapper around each model in my cross-validation metodology and generate SHAP values for each one.

I'm intrigued with the "look ahead" problem in time series forecast, and how it applies to SHAP explanaitions. Should I explain the the past based on the past itself with a model that knows it, or should I explain always the next month and aggregate those explanations?

Any clues or papers on how to create the metodology for appliyng SHAP in this cenario?

I didn't find any cases similar to mine.

Thanks, Lorenzo.