File 001 · partially declassified
Most trading ideas are noise.
We built the machine that proves it.
Vector Harness takes a trading idea (yours, a textbook's, an AI's), replays it against real market data, and keeps an honest score. We've been running it in private. Here's what we can show you.
Then we changed the setup: 4 trade lengths × 2 stop widths × 4 targets
Exhibit A · the scoreboard
Every popular rule lost money.
We took seven rules traders swear by and let each one trade every market for six years: one-hour trades with a stop and a target 1.5× as far, real spreads and slippage, never two trades overlapping.
Three of them (breakout, trend, momentum) do come out slightly ahead before costs. Costs wipe that out: after costs they lost in every year but 2022, when the breakout only broke even. One row is still sealed. We'll publish its score at launch, whether it wins or loses.
R = the amount risked per trade. 11,349 to 126,062 trades per rule. Every 95% interval is below zero, even resampling whole days (same-day trades move together) and correcting for testing 7 rules at once.
How we tested it, and where it could be wrong
- Data
- Every regular-session 5-minute bar on Alpaca's free IEX feed, Aug 2020 to Nov 2026. IEX is one exchange, about 2–3% of US volume, so its prices track the full market closely but its bars can miss a brief spike. The rules only use bars up to the decision; the trade only uses bars after it.
- Costs
- No commission. 1 bp (0.01%) when the target fills, 4 bp when the stop fills, 2 bp when time runs out, per round trip. That's tight for SPY and on the cheap side for single stocks. The three rules with an edge before costs would need costs 65–77% lower to break even.
- Overlap
- A rule never holds two trades at once on the same stock. Trades on different stocks the same day still move together, so the intervals resample whole trading days, not single trades.
- Many tries
- 224 setups is many chances to get lucky. None was profitable even before correcting for that, and the best setup per rule was picked on 2020–23 and checked on 2024–26. 2024–26 wasn't used to pick any setup, but it isn't fully untouched: we had already seen those years for the 1-hour setup. Live paper trading is the clean test.
- Stop and target in one bar
- Counted as a stop. It happens in under 0.1% of trades; counting it as a target changes no result by more than 0.002 R.
- Not tested
- Holding overnight, longer than half a day, tuned versions of each rule, and any AI model. That last one is the sealed row.
How it works
█████ an idea. Then try to break it.
- 01
█████
Short ███████, one trading idea each, every term █████████. ██████ 01: why most backtests lie. ██████ 02: ███ ███ ████ ██████.
- 02
Test
Every ██████ ends in an experiment, not a ████. Run it, or type your own rule. It trades six years of real bars, with costs.
- 03
Ask
A █████ walks you through why it won or lost, and ██ ██████ ████ ███ ████.
- 04
Verdict
Edge, coincidence, or eaten by costs. Then ██████ ███ ████.
What we're building
Three things. We'll tell you one.
Join, and we'll tell you two.
-
01
The Lab
Type an idea like "buy the first breakout" and the harness replays it on real bars, with costs, and tells you if it's an edge or a coincidence.
-
02
The ████
████ ██████ ████████ ███ ██████ ███████ ██ ████ ███ ██ ███ ████.
Waitlist only · join to unlock -
03
The ████████
██ ██ ███ ██████ █████ ███████ ███ ████ ████ ██ ████████ ██ █████.
Unlocks at launch
The sealed row opens for the waitlist first.
No spam. One email when it's ready, and maybe one before that if the sealed row surprises us.