Look-Ahead and Survivorship Bias
Look-ahead bias uses information before it was available. Survivorship bias excludes instruments that disappeared. Both can make an unprofitable historical rule appear stronger than it could have been in real time.
How it works
Use point-in-time membership, publication timestamps and delisted securities where relevant. Adjusted datasets can embed later corporate actions, and financial statements may be revised. Confirmed pivots need future bars, so the detection date differs from the pivot’s plotted date.
Worked example
A strategy tests today’s largest 50 companies back through ten years. Companies that failed and left the group are absent, while later winners are included before they qualified. Separately, using a quarter’s earnings release at the quarter-end date gives the system results that might not be published for weeks. Both need point-in-time corrections.
Amounts, prices and scenarios are fictional teaching examples. They exclude any effects not explicitly stated and are not trade recommendations.
Limits and common mistakes
Chronological splitting cannot repair contaminated inputs. Trying many rules and reporting only the best also creates selection bias. Keep a research log and count discarded experiments as part of the evidence.
Knowledge check
Does an out-of-sample split automatically eliminate look-ahead?
Read the explanation
No. If either training or test data contains future information, splitting the rows by date does not make those inputs historically available.
References and further reading
These sources explain the underlying concepts and specifications. They do not endorse QBMFX Academy. Check the relevant market, protocol version and reporting rules before applying an example.