Everything here failed.
We published it anyway.

The setup every COT account sells: positioning hits an extreme, so bet against the crowd.

metric: expectancy per trade, in R (risk units), 0.10% costs

50% hit rate, profit factor 0.91, negative in 17 of 21 years, and statistically identical in both eras. The point estimate sits inside its own noise band, so we do not claim it proves a loss. What it does show is that there is no directional edge at this granularity, in either direction, and costs eat a coin flip. Fading and following are both flat.

research/strategy.py

The confluence doctrine: do not catch the falling knife, wait until the reversal confirms.

metric: expectancy per trade, in R

This is the most useful result on the page, and it inverts what every confluence course teaches: the more confirmation you demand, the WORSE it gets, monotonically. Each added layer costs roughly 0.02R. Later entry surrenders move while stop noise stays constant. Confirmation buys comfort, not edge, and the data prices the comfort.

research/strategy2.py

The most-taught discretionary entry there is, tested standalone and then inside a COT extreme.

metric: mean 10-session ATR-scaled return · the small-n rows below the first are shown because we registered them, not because they conclude anything

Negative in both eras standalone, across 1,075 events. Over the same confirmation window the identical entry dates carried a +0.31R unconditional drift, so the zone logic did worse than doing nothing at all. The COT extreme does not rescue the zone. The two variants measured inside a COT extreme are n=56 and n=28: too few to conclude from, and shown only because we registered them in advance.

research/sd_cot.py

The obvious next move: only take the trade when the conditions that historically accompanied turns are present.

metric: mean R per trade, stop = max(median shakeout, 1 ATR)

Read the gated variant twice, because it is the whole argument of this site. Gating on the historically reliable conditions produced +0.20R in the years the rule had already seen, and nothing out of sample. A positive result in-sample that vanishes out-of-sample is the textbook signature of overfitting, and it is the exact pattern our methodology exists to catch. It also stopped out about 81% of its trades. We tested our own idea and it failed, so we are not selling it.

research/protocol_test.py

Trade a market's seasonal tendency: buy the months it has historically risen, sell the ones it has historically fallen · and stack it on a COT extreme for confirmation.

metric: mean one-month return, and the same return benchmarked against simply holding

Seasonality looks profitable and is not. The timed rule is long about 61% of months and these markets drifted up, so the positive figures are drift wearing a calendar: benchmarked against simply holding every eligible month long, seasonal timing LOSES in both windows (−0.02 pts train, −0.19 pts confirm). Stacked on a COT extreme it looks better · fades did 1.32 points better when the season agreed · but that gap fails in training and clears out of sample by a hair (interval lower bound +0.02%), which is one window of noise, not a finding. We also disclose a defect in our own bar: the adoption threshold was mis-scaled by a factor of twenty and the bootstrap interval it called for was never computed. Both were fixed after we had seen the numbers, so this study is exploratory rather than pre-registered, and we say so instead of presenting a repaired bar as the original.

research/seasonality.py

Follow the commercials: when their positioning is at an extreme of its 3-year range, trade with them.

metric: hit rate vs the same-direction baseline drift (49.9%)

A coin flip, and we are explicitly NOT claiming a kill. The confirm-era 8-week reading is nominally above the 49.9% baseline, the edges do not point consistently across era and horizon, and no significance test was registered for this study, so none is claimed. Two honest caveats: average signed returns turn negative in the confirmation era while the hit rate rises, and Williams himself presented this as a condition to be paired with price triggers, never as a standalone timing rule. We tested it as a timing rule, which is not quite what he sold.

research/williams.py