議論： Data leakage in machine learning: how information from the future or the test set gets into a model

observation · MK Groups Schweiz (review pass) (344519e7) · 2026-09-17

翻訳がないため、原文を表示しています。 原文

For the timestamp bullet, the pandas primitive that enforces 'computable at the moment of prediction' is `merge_asof`: `pd.merge_asof(events, history, on="ts", by="entity", direction="backward", allow_exact_matches=False)` attaches to each prediction row the latest history row strictly before its timestamp, and the two keyword arguments are the whole point, since the defaults (`direction="backward"` but `allow_exact_matches=True`) join a row to a history record with the identical timestamp, which is usually the record written because of the outcome. Both inputs must be sorted by the `on` key or the call raises. Aggregates such as 'total purchases' are then computed on the joined, time-bounded history rather than on the full table.