Your idea: I built a new swing model for gold to diversify my existing trades. It…

I built a new swing model for gold to diversify my existing trades. It's a mean-reversion strategy at swing highs and lows on XAU/USD. Entries run on the 1-minute chart. Not HFT, and that is the point. At this speed, I can run it on cheap infrastructure and leave it alone. Entry conditions: Price within 0.3x ATR of a swing level Price holding the right side of that level EMA 9 and EMA 21 aligned with the direction Enough volatility to be worth trading Backtest, May 2025 to July 2026: → 4,521 trades → 72.7% win rate → +81% gross That figure is gross on purpose. I keep costs out of the backtest because they change with the broker, the account, and the time of day, and if I bake one cost assumption into the model, I stop seeing whether the signal itself works. Gross tells me if there is an edge. Costs tell me whether I can keep it. So here is the second number: at roughly 0.01% per trade puts the realistic figure near 36%. The edge per trade is small, and it survives on trade count, which means execution quality decides this strategy more than the signal does. First five days live: → 12% return → 65% win rate Five days is not a sample. The live win rate is already below the model, and I expect it to keep moving toward the backtest number rather than away from it. The next input I want is news sentiment. I have tried this before and always hit the same wall: running headlines through a hosted LLM costs more in API credits than the edge is worth. I have now found an open-source model that looks promising and runs on my own hardware, so the cost problem mostly goes away. It still needs proper testing before it goes near live trades. Next step is more live data and a proper test on the sentiment input. I'll post what it does either way. I've made model weights public before. If this one turns out to be worth anything, happy to do the same here.