AC + MACD trend regime on BTC and ETH 1h

The idea, in plain English:

• Go long when Bill Williams' Accelerator Oscillator crosses zero in the same direction as MACD.
• Only trade with the trend: price above its 200-period EMA.
• Exits come from a parameter search: a stepped take-profit, a wide stop, and a trailing stop that lets winners run.

The Quant Agent drafted the code in a build run on 20 Sep. This is the kind of result that makes an agent look good, which is why we are publishing it.

How we tested it

The build run kept a locked hold-out. The agent only ever saw the training window.

• Pairs: BTC/USDT and ETH/USDT, 1h, spot
• Train: 20 Sep 2025 – 20 Jun 2026
• Hold-out: 20 Jun 2026 – 20 Sep 2026
• Eligibility: at least 126 trades for this window, and profitable on train.

What it did

• Train: 74 trades, +6.14%, profit factor 1.32, max drawdown 5.4%
• Hold-out: 31 trades, +7.25%, profit factor 2.34, max drawdown 5.4%

It made money on both windows. It also made 74 trades where the floor is 126, so the lab did not promote it. Asked later to add short trades, the agent found nothing better (66 trades, +2.58% on train).

The +42% we did not trust

Edge Lab shows this strategy as PROVEN, with +42.5% from January 2024 to September 2026 and 310 trades. Here is how that number was made:

• On 24 Sep we saved the agent's code as its own strategy, changed its config three times in about 15 minutes, and ran nine backtests the same day on overlapping windows.
• Those nine results ranged from −13.5% to +42.5%. The card shows the best one.
• None of them kept a hold-out back.
• Over the nine months the agent trained on, the version behind the +42% made +3.87% on 57 trades.

PROVEN in Edge Lab means the backtest passed our integrity checks, such as not peeking at future candles. It says the backtest is honest. It does not say the edge is real. We are changing the label so it also needs a locked hold-out and the trade floor.

Why it needs work

The numbers lean the right way: positive on train, positive on a hold-out the agent never saw. But 74 and 31 trades are too few to separate an edge from a good year, and the headline figure is the best of nine tries. Picking the winner of nine makes the winner look better than it is.

Two related versions are paper trading on Kraken futures: the original version (13 trades, +1.8% since 25 Aug) and a long/short variant (7 trades, +2.5% since 26 Sep). Neither is this exact strategy, and neither has enough trades to say anything yet.

What would move it forward:

• More trades: more pairs, or a longer window, with one hold-out set aside before any tuning.
• Paper trading this exact version, on the venue it was tested on.

What this teaches

• PROVEN should mean the edge survived data it never saw, not just that the backtest is honest.
• The best of nine backtests is not a result. Report how many you ran.
• A good hold-out on 31 trades is encouraging, not proof.
• Paper trade the version you tested, where you tested it.

What we did not claim

Not live. Not a return. Backtest figures are simulated on historical data, and paper figures are as of 5 Oct 2026.

The verdict is the product: promising, too thin to trust. Needs work.

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