How I sentence my own strategies to death

Eight product lines have been through the same rule here. Four are dead, four are unproven, none passed. The threshold is t* = 2.0, written down before any of them ran, and the thing being judged is not the level of t but its direction as the sample grows: a real signal pushes t up as data accumulates, noise pushes it down. The futures line was overturned that way on 26 August 2026, when t fell from 1.52 to 0.79 while n went from 30 to 53. It overturned a conclusion I had published on 20 August 2026.

That is the shortest description of what this site is. Everything below is how the rule works, and what it is not allowed to touch.

The threshold is written down before the test, not after

A threshold chosen after seeing the result is not a threshold. It is a description of the result. So t* is fixed in advance and it does not move: 2.0, against the 3.0 that Harvey, Liu and Zhu (2016) argue is the honest bar once you count how many things were tried. 2.0 is the loose line, and eight lines failed to clear the loose line.

The direction test matters more than the level. A t of 1.9 that has been climbing for three months and a t of 1.9 that has been sliding for three months are not the same object, and only one of them survives another month of data. Reading the level alone is how a fading line keeps its funding.

Counting the tests, not the winners

The multiple-testing count is kept because it decides whether any of this means anything. Every grid cell, every variant, every re-run on a different window is a test, and the count includes the ones that were abandoned. A strategy that looks significant after forty silent attempts looks significant because forty attempts were made. Bailey and López de Prado (2014) put a number on how fast that inflation happens.

What gets counted	Why
Every variant that ran	An abandoned attempt still used a draw from the distribution
Every window re-run	A new window is a new test, not the same test
Every line, including dead ones	Dropping the dead ones is how a survivorship record gets built

The arithmetic is unforgiving and it is supposed to be. A record that counts only survivors says nothing about the process that produced them.

The sample stopped growing? That is a broken pipeline

For 26 days the collector wrote nothing and every gate stayed green, because a series that stops growing looks exactly like a series with no new signals. That gap ran into 27 August 2026, and the fix was not a better alarm on the strategy side. It was to make each series print its own last date, so a stale one has to say so on the page. The performance page carries that per-series date, and the reasoning is in the piece on checks that pass for the wrong reason.

Backtests are assumed false until audited

A backtest is a claim about a computation, not about money. The audit that follows is where most of the claimed edge disappears: one line went from a claimed 13.25x to 2.26x once costs, fill assumptions and look-ahead were taken out. The method is in the backtest audit, and the distribution work behind the exit rules leans on the NIST/SEMATECH handbook rather than on a rule of thumb.

Reconciliation is the only evidence that does not depend on statistics

Every dollar figure here is reconciled against the account rather than recomputed from the strategy’s own records. The capital at risk is $3,300 and the broker either agrees with the balance or it does not. That is the one number nobody has to take my word for. How to verify a track record is the checklist a reader can run against this site, written so that some of its rows fail here.

What is published, and what never is

Published: the methodology, the decision rule, the multiple-testing count, how t* was chosen, the falsified conclusions, incident root causes, the list of data defects, every closed trade, net asset value and real dollar amounts, and the architecture and engineering concepts of the execution system (scanning separated from execution, a guard layer, dry-run mode, heartbeat and alerting, reconciliation logic).

Never published: entry and exit parameters and thresholds, holding time, stop placement, the capacity ceiling, the name of anything still open, and any live signal.

The line is one sentence: You can copy the engineering. You cannot copy the parameters.

This is not commercial secrecy. Three of the eight lines are counter-trend, so anyone following them raises the entry cost for everyone including me. The crypto line exits on a volume climax, and follow-on volume is itself a fake signal for that rule. And the argument this site makes, that I will kill my own ideas in public, stops being worth anything the moment somebody loses money copying it.

A build gate enforces the line: prose or template code naming a parameter, an account size, a capacity ceiling or an open position fails the build. It is a line-by-line string check, so it stops a slip and not an inference. Two individually harmless numbers in different paragraphs can imply a third, and no regular expression sees that. What stops it is a person noticing while writing, which means that layer is human, and humans get tired.

Frequently asked

Can I trade these strategies? No, and that is by construction. The parameters, thresholds, holding times and stop placement are not published. What is published is the decision rule and the falsification, and copying those harms nobody.

How was t* = 2.0 chosen? Written down before the runs and not touched afterwards. The literature argues for 3.0, so 2.0 is already the generous line, and eight lines failed to clear the generous line.

Why publish the things that failed? Because a site that publishes only what passed gives a reader no way to tell skill from survivorship. A record of dead strategies carries more information than a screenshot of a good month, and it can be checked. The coin data and the delisted list are there for that.

Is the Chinese version a translation of this page? No. The two versions share the numbers, the facts, and word for word the disclosure boundary above, because that boundary is a rule rather than an essay. The arguments and examples around it were written separately for each language. The Chinese original is 方法論：我用什麼判自己的策略死刑.