I have panel data consisting of yearly credit ratings as a target variable and some features for its estimation. Each year of my 20 year time series i have around 400 firms. I use shap to analyse some of those features and analyse how this results change over time.
I dont want to predict something i just want to analyse patterns. Do i need to consider overfitting? DO i need train test split? Some of my credit rating classes have only very few observations.
What would be the vest way? Train test split wouldnt be an option right?
