The half-split test

The problem it solves

Sweep enough data and you will find patterns. Most are one good week wearing the costume of a rule. The half-split test is the cheapest way Apex knows to tell them apart.

The test

Cut the record in two by date — not by row count, so that one heavy day cannot sit in both halves and test itself. Measure the effect in each half against the unconditional baseline. A finding passes only if the sign holds in both.

What it caught

• A "36.8 % zero-move" pattern that lived entirely on one side of a bug fix — an artefact, not a finding.
• A dollar pattern against Brent (−0.41 % the day after, 70 episodes) that halved between halves and later vanished on two years of data.
• A "bull run breadth" predictor whose sign flipped between halves — a coin toss.

What it does not do

It does not correct for multiple testing. Apex logs every hypothesis ever tested and judges significance against the whole family (a Bonferroni correction): when 276 tests produced 27 with p < 0.05, that was roughly what chance alone produces, and one finding survived.

See also: methodology · what is decision intelligence.