Data availability and look-ahead bias

Understand data availability and look-ahead bias through definitions, interpretation and the assumptions used in KabuForge research examples.

What it means

Look-ahead bias occurs when an evaluation uses information unknown at the decision time. Observation dates and availability timestamps can differ. Downloading history today or using revised values does not establish what was known then.

How to interpret it

Point-in-time checks compare decision time with available_at. Declared timestamp checks alone cannot certify vendor delivery, revisions, survivorship, corporate actions or the historical security universe.

Check the KabuForge example

The reference course uses prior-observation data and separates forward labels, but all runs remain RESEARCH-ONLY with pit_guarantee=false. Inspect input identity and each run’s receipt, then interpret results within the verified scope.

Separate observation dates from release times

Information released after today’s close cannot be used for today’s opening decision. An observation date matching the trading date does not make information available before its release.

Availability timestamps and survivorship

available_at

This timestamp declares when a record can be used for a decision. A later timestamp makes it ineligible for that decision. The declaration alone does not prove the vendor’s historical release record.

Survivorship bias

Backtesting only securities that exist today can omit securities that existed earlier but were later delisted. Retain the decision-time universe instead of treating the present universe as historical fact.