Can a neural network find an edge on gold H1? A walk-forward test with the answer fixed in advance
I wanted to know whether a neural network can find a tradable edge on gold H1 that simple rules cannot. Before running anything I wrote down the pass criteria, then ran a walk-forward test. The answer was no. Here is how it was done, so that you can check it or repeat it.
Data and trade rules
- HistData.com XAUUSD M1 (bid), 2010 to 2025, aggregated to H1.
- Entry at the close of a closed H1 bar. Stop 1.5 x ATR(14), target 3.0 x ATR (1:2), at most 72 hours, a bar that touches both counts as a stop. Cost 0.17 USD per trade.
- Features (no future information): returns over several horizons divided by ATR, distance to three moving averages, ATR ratios, position in the 24 and 120 bar range, hour and weekday, candle body and wicks.
Walk-forward
For each test year from 2014 to 2025 the models were trained only on the years before it (the last 5 days removed, because trade labels overlap). The test year was never used to tune anything. A model trades a direction only if its predicted expected value is above +0.15R, one position at a time. Models: LightGBM and a small neural network (two hidden layers, three random seeds averaged). Controls: the same models trained on shuffled labels, a rule that buys every bar, and a rule that buys above the 200 bar average.
Pass criteria, fixed in advance
- At least 200 trades.
- Mean of at least +0.10R per trade.
- The 95% block-bootstrap confidence interval (by month) above zero.
- At least 8 of 12 years positive.
- Better than the shuffled-label control by 0.05R.
- Better than the best simple baseline by 0.05R.
Result
| Method | Trades | Mean per trade | 95% interval | Positive years |
|---|---|---|---|---|
| LightGBM | 3,753 | -0.029R | -0.074 to +0.017 | 7 of 12 |
| Neural network | 2,010 | +0.047R | -0.014 to +0.110 | 7 of 12 |
| Buy every bar | 5,725 | +0.020R | 7 of 12 | |
| Buy above the 200 bar average | 3,413 | +0.045R | 7 of 12 |
Neither model passed. The neural network made +0.047R per trade, the same as the simplest trend rule (+0.045R), and its interval includes zero. It was strong in 2014 to 2017 (+132R in total) and slightly negative from 2018 to 2025 (-39R). That split was not part of the criteria, so I do not read much into it.
What I take from it
- Recognising a pattern in past prices is easy for a neural network. Finding a pattern that still pays after costs, in data that were not used for training, is the hard part.
- Fixing the criteria before the run matters: with a free choice of settings afterwards, some combination will look good.
- This is one setup on one symbol and timeframe. Other targets, features or timeframes may behave differently. It is not a statement that no edge exists.
Past data only; nothing here is a forecast or advice.
