The Register of Everything That Failed

Almost nobody publishes the strategies that did not work.

That is not an accident of temperament. A strategy vendor who published their failures would be publishing their inventory. A trader who publishes failures gets no engagement. An academic who publishes a null result often cannot get it into a journal at all — the file-drawer problem has been documented since the 1970s.

The result is a public record made almost entirely of winners. And if the only strategies you ever see are the ones that worked, you will badly misjudge how often anything works at all.

So here is our file drawer, opened.

The register

Every idea below was tested with the same pipeline: real broker spreads, real overnight financing, out-of-sample validation on data the rule never saw, and a significance bar fixed before the test ran.

Off-the-shelf products

Our own builds

Ideas tested without building an EA

The five ways a dead strategy looks alive

The register is the evidence. This is the part that transfers.

1. The sample is far too small to prove anything

This is the big one, and it is almost universal in trading marketing.

We measured our own testing engine to find out what it can actually see. The answer: it needs 1,476 trades to reliably detect an edge of 0.10R, and 5,903 trades for 0.05R. Below about 0.12R it is blind regardless of sample size.

A realistic good retail edge is 0.02–0.10R per trade. So a screenshot showing "+203% over 67 trades" is not weak evidence. It is no evidence — that sample cannot distinguish a real edge from luck in either direction.

Whenever you see a strategy sold on a track record, the first question is not "what was the return" but "how many trades, and is that enough to tell?" Usually it is not.

2. Drift gets mistaken for skill

Several gold moving-average rules are genuinely profitable after costs. They earn it by sitting long through 60–70% of an eight-year bull market and giving back part of the rise.

Score them against zero and they look clever. Score them against buying gold and doing nothing and all nine lose.

We measured this precisely on the EMA crossover: on equity indices, 60–72% of the apparent edge was simple directional bias — the signal was long most of the time in a market that rose. On metals it was only 3–11%, which is why metals looked different. That distinction is invisible in a backtest curve.

3. Correlated instruments fake breadth

"It works on 9 of 10 symbols" sounds like ten pieces of evidence. If those symbols are Germany 40, UK 100 and the Nasdaq, it is closer to one piece of evidence wearing three hats.

We watched a 9-of-10 result collapse to 4-of-10 out of sample for exactly this reason. In this month's indicator batch, the only signal that passed did so on three equity indices plus Brent — which is two independent observations, not nine.

4. The optimum is a search artifact

Our Donchian breakout test scored +29.21%/yr on silver. That number is true, and it is meaningless: it is the maximum of a 24-cell parameter search in a surface whose median was −4.04%/yr, and every instrument's best cell sat somewhere different.

A real effect clusters. Noise scatters. Report the surface, never the cell — the single best cell is exactly what an advert screenshot is made of.

5. Cost is compared to the wrong thing

The subtlest one, and it caught us this month.

We screened a gold scalper's costs by comparing the spread to the size of the move we hoped to catch, and concluded the hurdle was a survivable 8.1%. But a strategy's risk is defined by its stop, not by its hope. Measured against the actual stop, the spread was 0.152R per trade — every trade began with 15% of its risk already lost, which is above the smallest edge our engine can even measure.

The screen said proceed. The arithmetic had already ruled it out.

What this does and does not say

We want to be precise, because the overclaim is tempting and would be wrong.

What the evidence supports:

• These specific strategies, on these instruments, over these periods, failed a pre-registered testing standard.
• At retail CFD costs, short-timeframe trading faces a cost hurdle that frequently exceeds the size of any edge that is plausibly there.
• Track records presented in trading marketing are usually far too short to support the claims made from them.
• Some documented effects are real but unreachable through a retail account — commodity trend-following is genuinely supported in the literature, and CFD financing costs more than it pays.

What it does not support:

• That markets are perfectly efficient, or that nobody makes money. Institutions with better data, faster execution and costs an order of magnitude lower are playing a different game on the same screen.
• That a strategy we rejected cannot work for someone else, at different costs, on a different horizon.
• Any recommendation about what you should do with your money. This is a record of experiments, not advice.

The one thing worth taking away

Twenty-two ideas. Thirty-eight off-the-shelf products. The two that survived our cemetery in Field Report Nº 002 remain the only things still standing, and we have been explicit that we do not yet know whether they will hold up live.

That ratio is the finding. Not "trading does not work" — but that the base rate of a tested idea surviving is very low, and every part of the industry's marketing is constructed to hide that base rate from you.

The most useful thing we built this year was not a strategy. It was the ability to say no quickly, cheaply, and with a number attached.

---

Inkatech builds software and data tools for South African businesses, and runs this research openly because published nulls are worth more than private wins. Nothing here is financial advice. If you are testing a strategy and want to know whether your sample can even support your conclusion, get in touch.