Overview

A five-day return observed every day produces many rows but far fewer independent market outcomes. A model trained on one row can indirectly see part of the next row's target, even when the row identifiers are different. This topic makes the sample's information interval explicit and uses that interval to build a defensible evaluation design.

The previous topic established which inputs a historical decision could use. This topic asks which completed outcomes a researcher may use to fit a model and how the search for a model changes the meaning of the final reported result. These are separate sources of contamination: correcting timestamps does not correct overlapping train/test outcomes, and correcting overlap does not erase repeated strategy selection.

By the end of this note you should be able to:

• construct fixed-horizon and event-based targets aligned with feasible trades;
• calculate label concurrency and average uniqueness;
• define training, validation and final testing roles before model fitting;
• remove overlapping outcome intervals and justify an embargo;
• record the trials that determine how much evidence a winning result contains.

1. Contents
2. The big picture