We Audited Our Own AI's Win Rate. It Was Wrong. Here's the Real Number.
By Marcus Webb · June 12, 2026 · 6 min read
For weeks, our site said "80.1% win rate." In June 2026 we audited our own evaluation code and discovered that number was wrong — inflated by a counting bug. The honest figure, under the strictest methodology we could design, is around 48%.
We could have quietly fixed the bug and moved on. Instead, we re-evaluated every signal we've ever sent, replaced every public claim, took down videos that quoted the old number, and published the full methodology. This article is the complete story — and a guide to auditing any signal service the same way.
The Bug: Counting Wins Out of Order
Our evaluation code checked whether a signal's price ever reached its target — and only afterwards whether it hit its stop-loss. No chronological ordering.
Think about what that means. A signal drops 3%, hits its stop-loss — a real trader is out, with a loss. Two weeks later the stock recovers and touches the original target. Our old code looked at the peak, saw "target reached," and recorded a win.
🚩 This is the most common inflation mechanism in the industry. It usually isn't even malicious — it's a lazy backtest. But it turns losing systems into "80% winners" on paper. Ask any signal service: "Do you count a trade that hit its stop-loss first as a loss, even if price later recovered?" Watch them squirm.
The Fix: Three Strict Rules
- Chronological evaluation. Each signal is replayed day by day. The first event decides the outcome — if the stop is hit first, it's a loss. Forever.
- Same-day ambiguity = loss. If price touched both stop and target on the same day and we can't prove the order, we count it against ourselves.
- A fixed 14-day window. A target hit on day 40 doesn't count. No "eventually it worked out."
Then we re-ran every single signal in our history through the new code, against real NYSE/NASDAQ daily prices.
The Numbers, Before and After
| Old (buggy) | Corrected (strict) |
|---|
| Win rate | 80.1% | ~48% |
| Published losses | 26 | 109 |
| Avg return / signal | +3.36% | +0.49% |
| Best win | AVGO +39.6% (counted after stop-out) | NVDA +12.78% in 11 days |
Yes — the corrected win rate is below 50%. And the system is still net-positive, because winners average about +4% while losses are cut early at about −2.7%. That's how real trading systems work: asymmetry, not fantasy win rates.
The live numbers are public and update automatically every trading day, straight from the evaluation engine — no human edits the results page. See
the full track record: every signal, every outcome, every date.
What We Did About the Old Claims
- Replaced every "80.1%" claim across our site, blog, and bios with the corrected figures
- Took down or made private the videos that quoted the old number out loud
- Rebuilt our AI confidence score, which the audit showed was uncalibrated, into a deterministic, data-anchored 38–62 band
- Published the methodology on the results page, permanently
How to Audit Any Signal Service in 5 Minutes
- Demand the complete history — every signal with date, entry, target, stop. Not screenshots. If they can't produce it: walk away.
- Ask the stop-loss-first question — "if price hits the stop and later recovers to target, how do you count it?" Anything but "a loss" is inflation.
- Check the time window — open-ended "it eventually hit target" counting makes any system look brilliant.
- Look for the losses — a 90% win rate with no visible losing trades means the losses were deleted.
- Check the math coherence — does the win rate actually match the published wins and losses? You'd be surprised how often it doesn't.
⚠️ Not financial advice. Past performance does not guarantee future results. Marcus Webb is an AI-generated brand persona, not a real person. All performance figures are generated automatically by our evaluation engine and update daily on the results page.