Digest 001 ended with two promising results and a sealed holdout. Both results are now refuted, the holdout is spent, and a preregistered test of the premises under retail technical analysis failed on all four counts.
Three tests. Zero survivors. This is what a working process looks like.
The scoreboard
Every claim below was frozen before its data was read. That is the only reason these results mean anything: each had exactly one chance to survive, and the decision rule was written down first. There was nothing left to negotiate with afterwards.
Call-side asymmetry in 0DTE credit spreads (Digest 001's headline) — REFUTED. Median -1.16% of width, 3 of 7 markets positive.
QQQ near-money put spread (Digest 001's one survivor) — REJECTED. Edge decayed roughly 90%, fails positive-both-halves.
Four premises behind the popular TradingView trend scripts — 0 of 4. All failed at validation.
One — the replication
The call side was a window, not a law
Digest 001 reported that six of seven configurations surviving all three splits were call spreads. It also said, in writing, that this was a flag rather than a finding: all twenty configurations were inspected before the pattern was named, which is textbook multiple comparisons. So it was preregistered and re-run on eight markets the original never touched.
Seven staged; XLF was dropped with a recorded reason (Friday-only expiries, penny-flat quotes). The frozen rule required a positive median difference AND at least three quarters of markets positive. It returned median -1.16% and 3 of 7.
SMH calls beat SMH puts by 7.57 points of width. COIN calls lost to COIN puts by 11.11. The side that wins is a property of the name, not of the side.
One thing did replicate, and it is the boring one: 6 of 7 replication markets had a net-positive put side. Not 7 — XLE puts ran -0.33%, and XLE is one of the three markets where calls "won", so that win came from puts being negative rather than calls being strong.
Two — the holdout
The survivor decayed ninety percent
Digest 001's single live candidate — sell the QQQ 0DTE near-money put credit spread on Monday and Tuesday when credit is rich — had a holdout sealed from February to August 2026. It was spent once.
discovery: +5.69% per trade, Sharpe 2.942, 221 trades — passed
validation: +4.82% per trade, Sharpe 3.337, 43 trades — passed
holdout: +0.52% per trade, Sharpe 0.263, 72 trades — REJECTED
Eight of nine gates still passed. It stayed net-positive, survived doubled costs, survived removal of its best day, and cleared every sample floor. It failed on one gate: positive_both_halves. One half of the holdout window is negative.
So it is not nonsense. It is a rule that read +5.7% and +4.8% in-sample and +0.52% out of it, with Sharpe falling from 3.34 to 0.26. It was mostly measuring the window it was found in.
Three — the retail canon
Fibonacci marks the worst place in the range to buy
TradingView's trend listing is not what the listicles describe. Of the eleven scripts listed, three are Fibonacci tools, two are drawn zones, two are channels or ribbons. There is no MACD and no ADX. So the test targeted the premises those tools share, frozen and committed before any data was staged.
All four failed at validation. Worse, in that window every one underperformed simply being long: the baseline printed -3.11 bps and the premises printed -36.24, -22.64, -7.18 and -9.82. The moving-average ribbon also lost to its own single-MA control in both splits, so stacking averages made things worse rather than better.
A retracement of R sits at range position 1-R, so the famous 0.618 retracement is the 0.3-0.4 decile. The preregistration required that decile to beat both its neighbours. It was the worst decile in the range.
A correction worth recording: an earlier draft of this placed 0.618 in the seventh decile. That is the wrong end of the range, and uncorrected it would have flipped the headline to "Fibonacci works". It was caught before publication because the annotation was checked against the definition rather than against intuition.
Why publish this
An engine that only reports its winners is not measuring anything. The value of a preregistered negative is that it costs something to produce and cannot be manufactured after the fact — the rule was fixed, the data was read once, and the result is whatever it is.
Four sealed holdouts remain unspent. When they are spent, the outcome gets published the same way, whichever direction it goes.
Hypothetical backtests on research data. Not investment advice.