Target encoding inside folds

You encode industry identifiers using average future returns. Where must that encoding be estimated?

Estimate the mapping using training labels only, with out-of-fold or ordered encodings for training rows so their own targets do not trivially enter their features. Apply a smoothed training mapping to validation rows and define a fallback for unseen categories. In time-series settings, ensure every encoded statistic uses labels already available at the decision date. Smoothing helps variance, not leakage.