Gaussian channel on BTC and ETH 1h

We wrote the rule down in full before testing it:

• A Gaussian filter on hl2: 8 one-pole stages, period 300, no band offset.
• Go long when the filter is rising, price closes above it, Stochastic RSI is at an extreme (K above 80 or below 20), and price is above the 200 EMA.
• Only enter during the New York session (09:30–16:00, Monday to Friday), and close everything at 15:55.
• Exit when price closes back under the filter, with a fixed stop two ATRs below the entry.
• Long-only spot, real fees.

The spec also asked for a second run with the same signals and no session filter (24/7), and said to stop if both runs had a profit factor below 0.8.

How we tested it

We ran it as a Validate job. The Quant Agent implemented the rule as literally as it could and did not tune it. It ran the two required backtests on the training window, then the result was checked once on a hold-out the run never saw.

• Pairs: BTC/USDT and ETH/USDT, 1h, spot
• Train: 18 Sep 2025 – 18 Jun 2026
• Hold-out: 18 Jun 2026 – 18 Sep 2026
• Eligibility: at least 100 trades on 1h, and profitable on the training window.

Why we killed it

Both runs lost about 20% on the training window:

• Session on: 53 trades, 7.5% win rate, −20.0%, profit factor 0.09
• 24/7: 70 trades, 11.4% win rate, −21.2%, profit factor 0.14

Neither reached the 100-trade floor, and both hit the stop condition written into the spec. Fees were not the problem: they were about 0.1% of the gross. The losses came from the stops. Our read: the filter is smoothed over 300 bars, so by the time it turns up, much of the move is gone, and a stop two ATRs below a late entry gets hit by ordinary hourly noise.

The hold-out made money: 34 trades, +4.65%. That does not save it. A quarter and 34 trades cannot overturn a year that lost 21%. The hold-out is a check, not a second chance.

We had run the same spec once before, on 18 Sep, and got the same answer: 48 and 65 trades, profit factor 0.07 and 0.18. The result belongs to the rule, not to one run.

What this teaches

• Write the stop condition before you test. This spec had one, and it was hit.
• A session filter that cuts trades without improving them is not an edge.
• A heavily smoothed trend line plus a tight stop buys late and sells on noise.
• One good quarter on the hold-out does not cancel a bad year on train.

Why an earlier card said KEPT

An earlier version of this card said KEPT, with a +333% figure. That number came from a loosely described version of the idea, tested on the training window only, with a max drawdown under 1%. It never saw a hold-out. A result that clean usually means the backtest is seeing data it should not, so we retracted it instead of publishing it.

What we did not claim

Not paper. Not live. Not a return. Figures are simulated on one training window plus a locked hold-out.

The verdict is the product: two runs, both below the floor, both losing. Killed.

Any figure shown in a teardown is backtested or simulated and is not indicative of future performance.