The Currency Times

Errata

Corrections to notes already published here, each with its date and the reason. Kept in public because a record that edits itself quietly carries no weight.

Date
Correction
Affects
28 Jul 2026
The persistence figure quoted for intraday variance was computed on raw returns rather than deseasonalised ones, so what it measured was the daily activity cycle rather than the process. The note has been rewritten to say so.
Note 019, Volatility
14 Jul 2026
The scaling exponent was given as a point estimate with no interval. On a sample of that length the estimator returns values above one half on series with no memory, so the number needed a range around it to mean anything.
Note 023, Long memory
2 Jul 2026
The trial count for an earlier screen was not recorded before the search, and it cannot be reconstructed afterwards, so the deflated figure is unavailable. The result has been set aside until the screen is rerun with the count logged.
Validation, unnumbered
19 Jun 2026
Cross-validation folds were not purged in the first implementation, so overlapping observations carried information across the train and test boundary. Everything produced before this date is being rerun under the corrected procedure.
Method, gate three
3 Jun 2026
The exit adjustment was applied at the wrong sign on a subset of records. Corrected and the affected series rebuilt from the immutable raw layer.
Survivorship, Data integrity

Which gate would have caught it

Every correction above traces back to one of the five gates on the Method page. Sorting them that way turns a list of embarrassments into a diagnostic.

Gate OneThe claim was not stated precisely enough1Gate TwoThe data could not have been known then1Gate ThreeThe test saw the answer1Gate FourNot discounted for the search2Gate FiveThe outcome is recorded either way0

Gate four accounts for two of the five, which is unsurprising, because discounting for the number of trials is the discipline that has to be applied before the search rather than after it. Gate five has produced no corrections and never will, since it is the gate that produces this page.

How corrections are handled

A note is never silently edited. Where a finding changes, the note is rewritten, the change is logged here with a date, and the earlier version remains identifiable by its revision number in the provenance line at the top of the note.

Where a result cannot be corrected, because the information needed to correct it no longer exists, it is set aside with the reason recorded rather than left standing under a caveat. Two entries above are of that kind.

The value of a correction log is that it is uncomfortable to keep. That is also the only reason anyone should believe the rest.

NextOriginshow this started→