First-hand · verified against the real thing
What ~140 documented trading experiments taught me (most of them failed)
In one line: Ninety-five crypto runs, nine forex runs, a strategy freeze, and a blind re-test campaign that overturned the freeze — the honest summary of a research project whose biggest findings are negative.
QuantLab has run roughly a hundred and forty documented experiments: 95 numbered crypto runs, 9 forex runs, and a post-freeze blind out-of-sample campaign numbered T1 through T34. Measured by the goal the project started with — find a robust trading edge — most of it failed. Measured by what it teaches, the failures are the yield. This is the summary the research journal writes by itself.
The shape of the hunt
The crypto arc (R001–R095) ran the full zoo: momentum and mean-reversion families, volatility-breakout variants, machine-learning condition filters, ensemble and portfolio constructions. The forex arc (F001–F009) tested whether anything transferred. Then the project did something unusual: on 2026-08-09 it declared a strategy freeze — no more tuning, the promising configs locked — and opened a separate branch to re-test the frozen conclusions blind on per-year out-of-sample data with costs, deliberately trying to break its own results.
The negative results are the headline
Five-minute crypto has no cost-surviving edge — demonstrated seven independent ways in R089–R095. Not “we didn't find one”: proven, repeatedly, with different methods, that costs eat everything. A whole family of “winning” runs was retracted — the R066–R072 results looked excellent until an exit-model audit showed the measurement itself was flawed; they were re-registered as a proxy artifact. Most edges refused to travel — strategies that worked on some symbols failed on unseen ones (R044, R049, R078): they were descriptions of their training data, not edges. And the freeze itself failed — the locked configuration that survived in-sample scrutiny lost after costs on the strict per-year blind re-test, which forced a better, humbler final config. Deriv synthetics, tested for completeness, behaved like the random walk they are marketed as.
What a hundred failures actually teach
That the market's cost structure is a predator tuned to exactly the kind of edge an amateur finds. That a result you cannot re-derive blind is a rumor you told yourself. That discipline is a feature you build into the pipeline — journals, freezes, blind branches — because willpower is not a system. And that the promotion vocabulary matters as much as the tests: RETRACTED is a verdict the project has used on itself, in writing, and the log is better for it.
The flip side — what did survive, and how it was validated — is the validation gauntlet every result had to pass. The single most educational incident, where a champion's logged profit quietly evaporated under a correct audit, is the lookahead-bias story.
Research note: QuantLab experiments are presented for educational and research purposes. Historical backtests and simulations do not guarantee future results and should not be interpreted as investment advice.
Sources
Next