Research
Everything measured here, newest first. Most of it is a null:
a thing the trading literature asserts, put next to a control that removes one
condition, and found not to differ from it. The nulls are the point — they
cost the same to produce as a positive result and nobody else publishes
them.
This page has a feed. Every entry below points at a
page that states its own method, its sample size, and what it could not
measure.
2026-08-21
Every moving average is one of a small number of constructions. This puts twelve of them side by side with the same four numbers each — lag, noise gain, peak gain and stopband leak — all computed from the coefficients rather than from any market, so no backtest can disagree with them. The comparison EMA is matched on lag rather than on period, which is the only way the columns mean anything: two filters both called 24 routinely sit different distances behind price. Nothing wins twice.
2026-08-21
Adaptive indicators estimate a dominant cycle each bar and retune to it. That premise is testable without writing an indicator, and against a null of the same returns shuffled it does not hold: on 26 years of hourly bars from three majors, no spectral peak in price stands above chance, and a period measured on one window says nothing about the next. The positive control is the useful half — in absolute returns a peak stands up at 24 bars at six to twenty-nine times the band, which is the trading day. Fed a real cycle these estimators find it loudly.
2026-08-21
Build something smooth and hopelessly late, then subtract a deeper copy of it. T3 carries a dial that runs from a plain triple cascade at 34.5 bars of lag to exactly zero — not nearly zero — and the page prints the bill at every step: noise gain, peak gain, and how far the line carries past the top of a move before it settles. Zero lag has been possible since 1998. It has never been free.
2026-08-20
A Fibonacci tool marks three depths on a pullback and says price is more likely to turn there. That is a claim about counting, so it can be checked by counting: 25,934 pullbacks across 15 markets and four timeframes, measured only on the half of each instrument's history the method never saw.
2026-08-19
Of all the pullbacks that ever got this far, what share went on to make a new extreme. 94,347 pullbacks since 1999 across 15 markets, split per instrument at the midpoint of its own history and checked on the half that was held back.
How to read any of it
Three things are true of every measurement on this site. It is computed
from a stated rule over price, so a second person can implement the rule and
disagree with the number. It is checked against a control, or against a
holdout, or both — a rate on its own means nothing without the rate the same
question gets when the condition is removed. And the sample size travels with
the figure, because a percentage without one is a decoration.
None of it is advice, and none of it says what happens when you trade with
it. A base rate describes what price did; what you do about that depends on a
rule, and the rule is yours.