VALIDATION DESIGN

Out-of-sample validation vs. walk-forward optimization.

A single OOS test evaluates one frozen specification once; walk-forward analysis repeats selection and evaluation across ordered windows. They answer different questions and are not interchangeable labels.

Published
Updated
Version
2.0
Stage
validate
Level
Intermediate
Reading time
7 minutes
Scope
Time-ordered model evaluation
PATH

Before you begin

Learning objectives

  • Explain: One holdout, one decision
  • Explain: Repeated windows
  • Explain: Where leakage enters
01

One holdout, one decision

A conventional OOS design separates a final period, freezes the candidate and evaluates it once. It is simple to audit but supplies only one market path.

02

Repeated windows

Walk-forward might select on January to March and test April, then select on February to April and test May. The combined result describes a repeatable selection process, not one permanently fixed parameter set.

03

Where leakage enters

Choosing window length, objective or constraints after comparing the final result leaks information into design.

  • Keep time order intact
  • Record window rules before the run
  • Use a final holdout for one release candidate
  • Do not select a workflow from its final test curve
CHECK

Common mistakes

  • Reading the result without the stated scope and assumptions.
  • Changing the rule after seeing an outcome while still calling the data unseen evidence.
DO

Trader Checklist

  • Keep time order intact
  • Record window rules before the run
  • Use a final holdout for one release candidate
  • Do not select a workflow from its final test curve
APPLY

Practice exercise

Choose one of your own trading examples and write one page of rules, evidence and stopping conditions using the principles in "Out-of-sample validation vs. walk-forward optimization.".

FALSIFY

What evidence would overturn the conclusion

The conclusion should be overturned or narrowed if a key assumption cannot be reproduced inside the supported scope, the control cannot be executed, or new out-of-sample evidence repeatedly contradicts it.

SOURCES
NEXT