Skip to content

Register · Check 0026 · CIPHER

Does transfer entropy separate causation from correlation at production scale?

Scale is verified at 236 million records; no single result has been written up. A published instrument-level finding would settle it.

Outcome

Unestablished

Published

19 August 2026

Scale is verified at 236 million records; no single result has been written up. A published instrument-level finding would settle it.

Unestablished — so there is no result to publish in full

The apparatus of a complete entry — measurements, an independence table, a reasoning chain — would dress a non-result as a result. This check is listed because the question was asked and the answer has not been reached, which is a state the register is required to be able to express. What is missing is not the write-up. It is the finding.

Check 0031 is the standard a complete entry is held to.

What is established today

  • The corpus is 236,734,006 rows of minute and daily bars across spot and futures, so the measurement runs at the scale the question is about. That is a fact about the instrument, not about the question.
  • The bias correction it depends on is built and deterministic: raw transfer entropy is non-zero even for unrelated series, and a closed-form shrinkage correction replaced an earlier one that averaged twenty random shuffles. What is NOT built is the validation stage that would audit the leakage defences — so the measurement is careful, and nothing here establishes what it would show.

What would move it to full publication

  1. 01One instrument-level result: a specific pair of series, at production scale, where transfer entropy separates a causal relation from a merely correlated one — and the same pair analysed by a second method that could have disagreed.
  2. 02A stated null case. A separation that never fails to appear is not a discriminator, so a pair where the measure correctly reports no relation is part of the evidence, not an afterthought.
  3. 03The surrogate procedure written down before the result is read, so the significance threshold is not chosen after seeing the answer.