A 93%-win-rate strategy, run over 25 years: what survived
A teardown. Every number below is the tool's saved output; the reading at the end is ours and is marked as such.
2026-09-07 · author-published source, link on request
The claim. In August 2026 a trader on X posted a TradingView backtest of a "KNN machine-learning SuperTrend" strategy: PLTR, 15-minute bars, July 2023 – August 2026, +52.06%, 93.48% win rate (43/46 trades), profit factor 3.065, max drawdown 18.95%. The author published the full Pine source. A skeptic replied that the numbers "don't exist across a broad spectrum of stocks and timeframes" and asked for the same signal over the last 20 years.
What we ran. The author's published code, ported line-by-line with every default kept — SuperTrend (ATR 28, factor 7.0) · a KNN gate (k=10, 1,000-bar learning window, stride 10, 0.9 confidence threshold, PCA-compressed features) · a trailing stop (1.0% activation, 0.5% offset) · 100% of equity per trade · 0.04% commission per side · fills at signal-bar close. Nothing tuned. The port was cross-checked against an independent bar-by-bar reimplementation (agreement to ~1e-16) and the engine's trade list was verified trade-for-trade against a local mirror.
On what. Our feed caps 15-minute history at 60 days, so the 20-year axis ran on daily bars (2001-09 → 2026-08, the feed's 25-year maximum) across SPY, QQQ, AAPL, MSFT, NVDA, AMZN, JPM, XOM and PLTR; the nearest-timeframe axis ran on 1-hour bars (2 years) for PLTR, SPY, NVDA and AAPL. Deeper intraday history exists from paid vendors — this teardown didn't buy it.
Results — 13 runs.
| Run | Trades | Win rate | Strategy | Buy & hold | Verdict flag |
|---|---|---|---|---|---|
| SPY · 1d · 2001–2026 | 16 | 93.8% | +11.34% | +607.64% | UNDERPERFORMS_HOLD · Read with caution |
| QQQ · 1d · 2001–2026 | 15 | 80.0% | +0.47% | +2,025.9% | UNDERPERFORMS_HOLD · Read with caution |
| AAPL · 1d · 2001–2026 | 11 | 90.9% | +9.38% | +103,506% | UNDERPERFORMS_HOLD · Read with caution |
| MSFT · 1d · 2001–2026 | 11 | 90.9% | −3.29% | +1,753.9% | UNDERPERFORMS_HOLD · Read with caution |
| NVDA · 1d · 2001–2026 | 7 | 85.7% | +8.31% | +66,821% | UNDERPERFORMS_HOLD · Likely unreliable |
| AMZN · 1d · 2001–2026 | 13 | 76.9% | −11.47% | +62,516% | UNDERPERFORMS_HOLD · Read with caution |
| JPM · 1d · 2001–2026 | 4 | 50.0% | −11.21% | +857.35% | UNDERPERFORMS_HOLD · Likely unreliable |
| XOM · 1d · 2001–2026 | 12 | 75.0% | +1.66% | +282.49% | UNDERPERFORMS_HOLD · Read with caution |
| PLTR · 1d · full history | 0 | — | 0% | +1,860.95% | NO_TRADES · Likely unreliable |
| PLTR · 1h · 2024–2026 | 7 | 71.4% | −10.61% | +496.13% | UNDERPERFORMS_HOLD · Likely unreliable |
| SPY · 1h · 2024–2026 | 4 | 50.0% | −10.75% | +37.66% | UNDERPERFORMS_HOLD · Likely unreliable |
| NVDA · 1h · 2024–2026 | 6 | 66.7% | +2.47% | +96.81% | UNDERPERFORMS_HOLD · Likely unreliable |
| AAPL · 1h · 2024–2026 | 5 | 80.0% | +2.89% | +40.63% | UNDERPERFORMS_HOLD · Likely unreliable |
The tool's own words.
- SPY daily, the best of the 13: “Underperforms buy & hold by 596.3% — it lost to doing nothing (hold: 607.64%).”
- PLTR hourly: “Dies at realistic costs — … no fee level can make it profitable.”
- PLTR daily: “No trades were made — the conditions never triggered, so there is nothing to evaluate (holding would have returned 1860.95%, but sitting out isn't 'beating' it).”
Our reading (not the tool's).
- The win rate is real and it is not evidence. 76–94% replicates almost everywhere because a 0.5% trailing stop books many small wins and lets the rare unstopped loss run (SPY: average win +1.35%, average loss −7.74%). Win rate here measures the exit, not the signal.
- The profit does not replicate. Thirteen runs, none beat holding. The best result anywhere is +11.34% — over twenty-five years.
- The strategy cannot fire on its own demo stock's daily chart. Two nested 1,000-bar normalisation windows mean ~2,000 bars of warm-up; PLTR's whole daily history is ~1,490 bars.
- What this does not show: whether the indicator repaints on TradingView (a runtime question we can't test from outside), and how the strategy behaves on 15-minute bars over long history (we don't have that data).
Limits, stated. Trailing-stop exits are evaluated at bar granularity (the original fills intrabar). The 4-hour EMA(50) filter was mapped by time horizon (an EMA spanning ~200 regular-trading hours) because our feed has no 25-year 4h series. Indicator warm-up seeds can differ from Pine's. None of these move a buy-and-hold gap of hundreds of percent.
Receipts
13 share pages, open without an account (re-issued 2026-09-07 after the share-page fix; each shows the tested window, bar count, verdict and — where the tool computes it — the cost-stress panel):
- SPY 1d: https://stlab.app/v/ecba0721-74bc-4851-8d6b-2554cf229aeb
- QQQ 1d: https://stlab.app/v/440bf4d6-c4d9-42b7-b0e5-8bed5474d566
- AAPL 1d: https://stlab.app/v/a2af23d7-cc2e-48fe-ae4c-e1867a010cc2
- MSFT 1d: https://stlab.app/v/0d74e892-2cb5-409a-ab70-1cb56d4a43fb
- NVDA 1d: https://stlab.app/v/5ae385c0-99ed-43c1-b3e2-10e205cec992
- AMZN 1d: https://stlab.app/v/73958a62-a721-43d7-a397-8f60107edcfe
- JPM 1d: https://stlab.app/v/7f597b8d-babb-486e-99a7-dcac89115e86
- XOM 1d: https://stlab.app/v/b8849c5b-1140-48b5-a3a0-b900df70062e
- PLTR 1d: https://stlab.app/v/14870552-fa4a-46ce-bc85-d186ad890e43
- PLTR 1h: https://stlab.app/v/98ecbada-a523-4aef-8e7e-472f7d041027
- SPY 1h: https://stlab.app/v/2ddd3b42-3e3f-4eb2-98b3-4483bbef9b6d
- NVDA 1h: https://stlab.app/v/5bc7ddf2-485d-4046-bf1f-9c409c979c5f
- AAPL 1h: https://stlab.app/v/ebdacd2f-bda4-4d23-abc3-0c9f79ab0056
Simulated results on historical data — research output, not financial advice. Past performance does not guarantee future results.