The lines that did not survive.

configurations tested

9

configurations that won

0

premium collected, best case

$39,000 on $5,900 of capital

ending value

$47,019

buy-and-hold ending value

$69,513

Closed August 2026 Nine of nine lose, by 0.31 to 7.26 points of annual return.

Three roots, three strike distances, fourteen years, every leg sold at the bid and settled at expiry. The most flattering configuration collected 660% of the account in premium and still finished $22,000 behind holding the index. That gap is the whole lesson: the income is real and large and completely beside the point, because it is the price received for the upside that was sold away. Reporting premium collected as if it were profit is how the strategy is marketed, and it is exactly what a comparison against the alternative exposes.

campaigns simulated

8,182 across 228 names

win rate, best ladder

99.8%

its annual return

0.111%

owning the same names

15.11%

index version

100.0% win rate, 0.041%/yr

Closed August 2026 Undefeated, and the worst available use of the capital.

Every rung added to the ladder raises the win rate and lowers the return, monotonically, in every configuration tested. On indices the strongest form of the defence is literally true — across 177 campaigns it never lost a single one, and its worst case was still positive — while returning 0.041% a year against 14.37% for holding the same index with the same money. The reserve is the cost: capital committed to funding rungs that mostly never get used is capital not invested. A strategy can have a perfect record and still be the worst thing you could have done.

max drawdown, S&P, signal

−34.1%

max drawdown, S&P, buy-and-hold

−34.1%

random books beating it, S&P

75% on return, 77% on return/drawdown

small caps

drawdown deepened, −42.3% to −51.9%

roots where it beat random timing

1 of 3

Closed August 2026 On the index everyone quotes, it delivered no crash protection at all.

The control is what decides this one. Any rule that sits in cash a quarter of the time shows a smaller drawdown whether or not it means anything, so the signal has to be measured against random timing matched on time-in-market, not against buy-and-hold. Against 200 such random books it lost on two roots out of three, and on small caps it was worse than useless — it deepened the drawdown while cutting the return by two thirds. The headline number is the S&P drawdown: identical to buy-and-hold, to the tenth of a point. The signal's entire selling point, on its most-cited market, is absent.

breakouts measured

12,954 at the shortest horizon

three-month return

+2.432% (t 10.23)

random days, same uptrend

+3.367%

difference

−0.935% (t −3.15)

horizons where it beat the control

0 of 3

Closed August 2026 Significantly worse than buying the same stock on an unremarkable day.

In isolation the arm looks superb: a t of 10.23 over three months is the sort of number a breakout course is built on. But a stock at a yearly high is in an uptrend by construction, so the only honest control is random days drawn from that same stock's uptrend, at the same firing rate and holding period. Those ordinary days returned more, at every horizon, with t between −3.15 and −4.10. The signal does not merely add nothing. It subtracts — you are paying attention to the chart in order to buy at a worse moment than if you had not looked.

entries, shared across arms

9,120

no stop, at matched volatility

14.39%

tightest stop, matched

3.35%

widest stop, matched

6.61%

worst case, no stop vs stopped

−80.5% against −82.9%

Closed August 2026 They cost more than they save, and they did not hold at the extreme either.

The two tests people usually run are both rigged. 'It reduced my drawdown' is trivially true, because a stop makes the position smaller in expectation — so does trading smaller. And grading in R is arithmetic rather than evidence: drift is fixed in percent, so drift measured in R scales as one over R and tighter stops win mechanically without predicting anything. Any study concluding that tighter stops have better expectancy in R has discovered division. Measured in percent on matched capital and then compared at matched volatility, every stop lost. The worst case got worse too, because gaps go through stops.

candidate pairs searched

~26,000

alpha, chosen honestly

−0.04% (t −0.08)

alpha, random pairs

−0.17% (t −0.11)

alpha, pairs chosen with hindsight

+2.73% (t 4.84)

Sharpe with hindsight

1.69

Closed August 2026 Zero — unless you pick the pairs knowing how they turned out.

Same rule, same threshold, same universe. The only thing that changes between the three rows is when the pairs were chosen. Formed on one window and traded on the next, the strategy returns nothing, at a t of −0.08 — not weak, not marginal, zero. Worse for the method: carefully selected pairs are statistically indistinguishable from randomly assembled ones. And chosen using the whole sample, the identical rule produces a Sharpe of 1.69 and an alpha at t = 4.84. That last row is the one that gets published, and the gap between it and the first row is the entire value of insisting selection happens before the tested window.

announcements traded

57 and 51 across two names

earnings arm

−0.345% per trade (t −2.75)

same option on a random day

−0.104%

win rate, earnings

36.8%

Closed August 2026 Worse than selling the identical option on an ordinary day.

The question that matters was never whether implied volatility falls after the announcement — it always does, that is arithmetic. It is whether the premium collected exceeds the move the event produces. It does not. Selling the put the session before and buying it back the session after lost more than three times what the same trade lost on matched non-earnings windows in the same name, at the same tenor, over the same holding length. The crush is real, it is priced, and what you are actually being paid for is standing in front of the announcement.

expirations tested

50

closes nearer max pain

50.0% (z +0.00)

mean miss, max pain

1.647%

mean miss, guessing the open

0.527%

expiries max pain won

16.0%

Closed August 2026 Three times worse than doing nothing.

Max pain was computed from the full open-interest ladder as of the prior session, so nothing about the outcome could leak into the prediction. Price finished nearer the max-pain strike than it started on exactly half of expirations, and the average distance to it grew through the session rather than shrinking. The decisive comparison is not against chance but against the laziest available alternative: predicting the day closes where it opened is three times more accurate. Fifty expirations is a modest sample, but a predictor that loses to guessing the open on 84% of occasions is not a marginal effect waiting for more data.

sample

2,130 sessions

beta to the S&P

0.96

alpha

+1.87%/yr

alpha, t

0.22

vs buy-and-hold

12.10% against 14.36%

Closed August 2026 It was beta.

The only version a cash account can actually trade moves one-for-one with the S&P and returns less than simply holding it: worse compound return, worse Sharpe, deeper drawdown, an alpha indistinguishable from zero. The premium is real, but it is payment for crash risk — earned through calm years and handed back in the volatility events — rather than a skill anyone is being compensated for. A second finding made the test itself awkward: the condition is true on 91.3% of sessions, which is not a condition, it is 'be invested'. A matched-random control drawn against a signal that is on 91% of the time selects nearly the same days and has almost no power to reject anything, so the control it had already passed meant nothing.

cells tested

~440 across four lanes

survivors

0

concentration

7% of rows held 84% of the effect

top ten names removed

t −0.46

against a market-signed benchmark

+0.026% (t +0.63)

Closed August 2026 There was no baseline left to condition on.

The effect everything was being conditioned against turned out to be a fat right tail in a handful of high-volatility names rather than a mean: removing 7% of the rows sent the result negative. Leave-one-out resampling, the standard check for exactly this, was blind to it and passed. The benchmark was wrong as well. Each pattern's outcome had been signed in that pattern's own favoured direction, so bullish and bearish rows cancelled inside the control and it removed only about a third of the market's own drift. Measured against a market-signed benchmark instead, the headline moved by a factor of 4.8 — to nothing.

the bar

15.85%/yr

average published effect

7.9%/yr gross, in-sample

net, post-publication, post-2005

0.96%/yr

cost of rebalancing monthly

1.08%/yr

162 anomalies, out of sample

+0.14%/mo before borrow, −0.01%/mo after

Closed August 2026 A level mismatch, not bad luck.

The average published anomaly delivers about half the required return before a single cost is charged, and every adjustment after that points downward. The decisive pair is the middle two figures: rebalancing monthly costs more than the average anomaly nets. Being small helps with exactly one of the four cost terms, price impact, and destroys the thing the effect depends on — a cross-sectional edge of a few basis points a month exists only as an average across hundreds of names, and held in the handful of positions a small account allows it is a coin flip with a thumb on the scale. Because the gap is arithmetic rather than statistical, reading further into the literature cannot close it. That is why this entry closes a source rather than a strategy.

median best t, 20 cells searched

2.16

median best t, 188 cells searched

2.69

best t available anywhere

3.319

its p-value

0.067

Closed August 2026 The ceiling of the search space is not significant.

The data was permuted to destroy any real effect and the same sweep re-run six hundred times. A best-of-search t around 2.5 is the median outcome of that — the ordinary result of searching noise, not evidence of anything. A t of 2.8 found after sixty attempts works out to roughly p = 0.17, which is unremarkable. And the single best result obtainable anywhere in the legal search space, on the real data rather than the permuted data, still does not clear significance. That closes the strategy of looking harder: the space does not contain a finding, so no amount of further search can produce one.