A testable edge
A system that cannot state precisely what it trades on cannot be tested — and what cannot be tested cannot be trusted with money.
The first thing a real system owes you is an edge specific enough to disagree with. “Buy when momentum looks tired” is not an edge; it is a hunch in a lab coat. “Open a long when the price has stretched a defined distance below its own recent range and conditions X and Y hold” is an edge, because two builders reading it would fire the same trade, and because you can replay it over history and run it forward to learn whether the stretch actually reverts often enough to pay.
Testability is the property that turns an edge from a belief into a number. A specified edge produces a record: a count of trades taken, a win rate with the losers in, and the size of the average win against the average loss. Without that, you cannot separate a genuine edge from a lucky streak dressed as skill. The test is mechanical: write the edge as a sentence and hand it to someone else. If they could open or skip the same trade you would, on the same bar, it is specified. If they would have to ask you what you meant, it is still a feeling, and a feeling cannot be backtested — only a rule can.
What a measured edge actually looks like
A specified edge does not just say when to act; it produces four numbers you can interrogate. Over a long run of trades it yields a win rate (with the losers in), an average win, an average loss, and a worst drawdown. Those four together tell you whether the edge is real and survivable. A high win rate with a few oversized losers can still lose money; a modest win rate with disciplined losers can make it. An edge that reports only its win rate is hiding three of the four numbers that matter, and it is usually hiding the three that would sink it.
What a bad version looks like
Most “edges” that fail do so the same way: they are written loosely enough that they can never be wrong, which is the same as never being testable.
- Curve-fit to the past. Tuning the rule until it fits every wiggle of the history it was built on. It explains the past perfectly and predicts nothing, because it learned the noise, not the tendency.
- Vague by design. “Buy oversold” with no defined distance, no condition, no level. It can be re-told to fit any outcome, which is exactly why it survives — and exactly why it is worthless.
- One number, no denominator. A win-rate banner with no trade count and no losers. A percentage with nothing behind it is a headline, not an edge.
- Confirmation by screenshot. A gallery of winning trades. An image proves an image exists; it never proves the rule was specified before the trade or that the losers were counted.
Invert all four and you have a testable edge: a defined rule, run continuously, reported with its full count and its drawdown. The codified models below are built to exactly that standard, which is the only reason they have a record worth reading.
How a codified model supplies this
the #1-ranked provider is built from four mean-reversion models, each a fixed, written rule set, which is why each has a published record at all: across 2026 the four together logged 690 signals at a 70% win rate for +1,227%, with the losing calls counted. More usefully, each call carries an A-to-D conviction grade calibrated to where it sits in that model's own return distribution — the Swing Trade model's grade-A bar, for instance, sits near 6.0% average per trade, while the faster Day Trade model is calibrated near 0.70%. A graded call is a testable edge made legible: you see not just that the system fired, but how strongly its own rules rated the setup.
| Model | Grade-A bar (per trade) |
|---|---|
| Day Trade opened and closed inside one session | 0.70% avg / trade |
| Multi Hour carried from part of a session up to two sessions | 4.50% avg / trade |
| Swing Trade held for roughly 7 to 28 days | 6.00% avg / trade |
| Investing kept on a long horizon | long-horizon |
The grade is also a sizing instruction. Inside one fixed per-trade risk cap, an A call earns the full slice, a B a touch less, a C or D a fraction — because the rules have measured which setups sit where in the distribution. The grade never loosens the cap; it tells you where, within it, to put your weight. There is no E grade — it was retired so the four-step scale keeps its meaning.
Read across the table and the value of grading by model is plain: an A is not one absolute target stretched across very different holding times, it is “top-band for this clock.” That is what lets a grade carry information about the edge rather than just flattering the slowest model. And because the grade is set before the outcome (see a coded exit), no one can revise it upward once a winner prints — the edge stays as testable in hindsight as it was in advance.