My Masterpiece · Built in 23 Days
It invented 9,300 trading strategies, ran 486,000 trials, and tried to kill every single one. The one that survived crushed buy-and-hold.
The Champion
15.2%
a year.
Out-of-sample, against 1.1% a year for buying and holding the sector’s five largest stocks, July 2024 to May 2026.
9,300
Strategy designs generated and screened
486,000
Parameter trials
1
Champion left standing
1,335
Automated tests
The Idea
A backtest that looks great is easy to build and usually means nothing. The hard part is knowing which results are real.
So I built a research machine that generates strategies by the thousand and then attacks every one of them. It runs the whole pipeline, from market data to strategy search to validation to live orders on a brokerage paper account, reconciled to the penny.
The Results
A breakout strategy on the energy sector. It buys when a stock breaks its 20-day high, holds up to 10 days, and protects itself with a trailing stop that adapts to volatility.
Out-of-sample test
15.2% /yr
Against 1.1% a year for buy-and-hold, on market data it was never tuned on. 170 trades, July 2024 to May 2026.
Harshest cost stress
2.5% /yr
Triple the slippage and per-share commissions together, and it still came out ahead of the benchmark.
First live paper run
+1.20%
Against +0.37% for the energy sector index fund, May 14 to 26, 2026. Every position reconciled to the broker to the penny.
The Gauntlet
Every strategy combines an entry signal, a filter, a way of combining them, an exit rule, and a market hypothesis. The system builds every combination, throws out the duplicates and the nonsense, and then makes each survivor prove itself again and again.
17 entry signals, 17 filters, 4 combiners, 5 exits, and 5 hypotheses make 28,900 combinations. De-duplicated and rule-checked, 9,300 remain.
Beat buying and holding, and clear a 12% annual return, over 2022 to 2024. Most designs die here.
250 tuned trials per strategy, scored across five separate time periods with a buffer between them, run in parallel across 23 processor cores. It has to win in most periods, keep its losses in check, and trade often enough to mean something.
Now prove it on market data it was never tuned on.
Still beat the benchmark with triple the slippage, per-share commissions, and both at once.
A final review of the nine survivors crowns the cleanest, most robust design.
Why the Number Holds Up
Signals are computed at the close and filled at the next open, and every calculation looks only backward. Date windows are checked automatically, so testing never leaks into research.
The live decision logic is checked against the backtest engine by automated tests, and again before every single order. If they disagree, nothing is submitted.
Seeded searches, fixed ordering, and frozen reference results mean any run repeats exactly. Any change in behavior gets caught.
A soft halt at 10% below peak, a hard halt at 20%, a daily loss limit, and a guard against every position riding the same bet.
A manual catch-up run risked counting one trading day twice. I traced it, corrected the positions, and added a guard that refuses to process any day twice.
The champion choice, its backup, and the reasons live in an append-only log. Every live day produces a reconciliation, a replay, and an audit file.
One Developer
68,000
Lines of code
25,800
Lines of tests
1,335
Automated tests
2,114
Commits
More test code than application code, because that’s what software that handles money deserves. AI tools sped up the typing; every research design, every gate, and every acceptance decision was mine.
Checks that verify themselves, records that reconcile, and nothing shipped on a hunch. The same discipline runs an insurance agency’s carrier audit, which re-checks every change before it writes a thing.
A private system I run locally on a brokerage paper account for my own research. It isn’t offered as a service or as investment advice, and the code stays private.