Look-ahead is the classic quantitative error: a value calculated at the end of a session gets copied onto its earlier bars, and a model learns from a future it will never see.
May 2026
NOVAQ traced one such case in a day-type component of a bias model. A small probe of 50 events found 6% disagreement between stored values and point-in-time recomputation. A full run over 5,010 events found 13.83%. An alignment effect halved on NQ, from +18.33 to +8.47 R a year, and turned negative on ES. Point-in-time columns became mandatory.
October 2026
A wider review then asked the larger question: what happens to the whole rule-based book when each input is rebuilt exactly as it was known at the time? Of 58 books, 44 did not change at all. The total fell by 8.4%.
Why that is a strong result
Most of the book never depended on the faulty inputs. The 14 books that did change had individual cuts of up to about 80%, and the corrected figures now form the base of all published portfolio results.
What we learned
Fixing an instance is not fixing a class. Each confirmed error now calls for a search for its relatives, and the best protection is a test that fails on any new use of the faulty source.
Limits
The review addresses this class of look-ahead.