subtopic

Data dictionary and leakage watch-list

Build a table: column, meaning, unit, type, allowed range, 'available at prediction time?' Yes/No. Concepts: Data dictionary; leakage-prone and post-outcome fields; identifiers Project A: Build the RENAME dict (long names → snake_case) and decode A-codes if you use UCI raw. Leakage: no post-outcome columns here (all are application-time). Mark PROTECTED attributes: personal_status_sex (sex + marital status), age, is_foreign_worker — used for fairness checks, not blindly as model features. Project B: Flag 'total load forecast' and 'price day ahead' (a TSO forecast and a market result: use only as a benchmark, not as features); flag every 'actual' generation column as unknown at forecast time.