Edge Autopsy邊際解剖

8 lines · 0 passed · $3,300 capital · 1 stale · 2026-08-27which one?

$3,300 of capital, 8 lines, none past the threshold.

One set of criteria, fixed in advance, applied to 8 markets. 4 killed, 4 unproven, none passing.

Where the 8 lines stand

Line t Against the t*=2.0 bar n Verdict
Altcoin spotSignal from perpetual funding and open interest, position held in spot 1.88 — unproven
The closest to the bar and still short of it. $1,848.42 live across 8/20 slots. Stress test: drop the ten best trades and the whole thing turns negative — the return is concentrated in a few, so the exit mechanism outweighs the entry price.
Taiwan short squeezeDirection after margin-suspension and short-to-margin events 1.67 — killed
All three variants dead: strongest cell t=1.67, weakest 0.05. No capital committed, data pages deferred.
US equity fadeMean reversion after a single-day spike 1.40 — unproven
I audited my own backtest: 13.25x came down to 2.26x. All three overstatements pushed the same way, none the other. $1,316.64 live.
Futures / CFD fadeFading order flow, four instruments across four sessions 0.79 53 unproven
A week earlier t=1.52 at n=30; five trading days later 0.79 at n=53. The shadow replay over the same window rose to 2.59 instead. The split was resolved on 2026-08-27: entry latency accounts for 82% of it and cost for 18%, and the latency accumulates in one direction within a window. The fix shipped on 2026-08-27 and is not yet accepted — acceptance reads latency in seconds, not profit.
Solana meme ladderShort-horizon rotation with laddered scaling — no test needed — killed
No test needed: six days of expectancy between -13% and -18%, not one day positive. $0 committed, data not retained.
Polymarket makingReward-driven market making, 20 days live before withdrawal — no test needed — killed
Every daily P&L was positive and the wallet barely moved: +$78.86 on paper against +$6.36 realised, the -$72.50 difference sitting in inventory. A 60-second markout on a book that trades single digits an hour is noise rather than profit — the whole statement was on the wrong time scale.
Sports divergenceDivergence between the de-vigged fair price and the prediction-market book, taken and held to settlement — no test needed 15 unproven
Never traded with real money, 15 paper entries. Theoretical net EV +2.09% against +12.45% realised — the gap is luck at n=15 rather than edge, and the former is what belongs in the ledger. Separating a 60% CLV hit rate from 50% needs n~152, roughly 100 days at this line’s signal density.
Trading-bot SaaSSell the tool instead of trading it — no test needed — killed
Dead on the commercial side: 33 payment attempts totalling $1,201, every one timed out and cancelled, revenue $0. Eight recorded technical failures as well, three of them the same bug class — a check passing for the wrong reason.

n is shown only where the sample is final; where it is not, the cell stays empty rather than carrying an estimate. The Taiwan line takes the strongest of its three variants — killing a line on its most favourable number is more honest than killing it on its worst.

There is nothing new in Wall Street. There can't be, because speculation is as old as the hills.

Edwin Lefèvre, Reminiscences of a Stock Operator, chapter one, 1923. The line is usually attributed to Jesse Livermore — the book is a roman à clef and its first-person narrator, Larry Livingston, is modelled on him. Lefèvre wrote it. Full text (Project Gutenberg), retrieved 2026-08-27.

The criterion: which way t moves

The threshold was fixed in advance at t* = 2.0, and what it judges is the direction rather than the level: add sample to a real signal and t rises, add sample to noise and t falls. All 8 panels share one pair of axes, so they compare directly.

Altcoin spot 1.88 One reading, no t-versus-n series
no track recorded TW squeeze 1.67 Monthly clustered t only, no series by sample size
US fade 1.40 One reading, no t-versus-n series
Futures fade 0.79 Live going down, shadow going up
no track recorded Solana ladder — No t test; expectancy was negative outright
no track recorded Polymarket — No t test; 20 days measured at +0.49%
no track recorded Sports divergence — No t test; CLV+ on 9 of 15, p = 0.3036, not significant
no track recorded Bot SaaS — Not a trading line, no test statistic

t runs 0 to 3 on the vertical axis and sample size n runs 20 to 64 on the horizontal, shared across all 8 panels; the red dashed line is the 2.0 threshold. 7 panels have no track — 2 hold a single reading, 1 have a test statistic but no series recomputed by sample size, and 4 were never t-tested (the dashed-outline panels). Empty stays empty; no line gets invented to complete the picture. The only track that crosses the threshold is the futures line's shadow replay, which holds no money at all. The tracks come from this note (in Chinese).

Notes

Contract data

The signal source behind the altcoin spot line, published in full: daily funding rate and open interest for 718 Binance perpetual contracts, one page each. 29 of them have been delisted — most market sites delete the page once a contract is gone, these stay, and that is the part of this data set which is hard to replace.

Index of all 718 (in Chinese) · The 29 delisted (in Chinese) · Last day in the database 2026-08-10. Not live quotes, and no signals. Where each contract's funding and open interest stop is printed on that contract's own page — no single date holds for all 718.

What is never published

Parameters, thresholds, holding times, stop placement, open position names and live signals are never published; method, falsification results and every closed trade are. The boundary and the reasoning behind it are on the method page, and a build gate enforces it.