We test far more than we trade. Every variant we reject is logged, so the three bots that run are the ones that survived, not the ones that looked best once.
$2.50 commission per side and one tick of slippage per side on every simulated trade.
Every neighbouring setting has to hold up. If a small change breaks it, it was luck.
A new bot has to add a different market or hour, not more of the same risk.
If the live setting were a lucky peak, the settings around it would fall apart. They do not.
| Bot | Setting | Trades | Profit factor | Net, 1 contract |
|---|---|---|---|---|
| LIVE TB48 | Stop 1.0 ATR · exit 11:00 | 121 | 2.47 | +$18,058 |
| TB48 | Stop 0.8 ATR | – | 1.89 | – |
| TB48 | Stop 1.2 ATR | – | 2.24 | – |
| TB48 | Exit 10:00 | – | 2.18 | – |
| TB48 | Exit 12:00 | – | 2.02 | – |
| LIVE TB34 | Threshold 0.55 · stop 0.35 | 159 | 2.53 | +$13,621 |
| TB34 | Threshold 0.45 | – | 2.52 | +$15,200 |
| TB34 | Threshold 0.65 | – | 2.60 | – |
| TB34 | Stop 0.25 | – | 2.41 | – |
| TB41 | Trend filter 0.05 to 0.15 | – | 1.83 to 1.87 | – |
The homepage shows the last three years, where the yen bot was quiet. This is its full 15-minute history from 2001 to 2026, so a slow stretch can be judged against two decades of data.
Hypothetical backtest, one contract, $2.50 commission and one tick of slippage per side. Not live trading.
Moving TB48's trigger from 0.65 to 0.60 ATR cut net from $25.6K to $21.7K and the last three years went flat. 0.65 stays.
Timed moves on days TB41 sits out: best profit factor 1.13 on 2019–2026. Mean reversion on the same days: 1.25. Neither ships.
Letting TB41 take a second long added $6 over 105 extra trades and dropped the profit factor from 1.91 to 1.80.
Three families, 7,422 configurations. Nothing reached the quality bar of the live gold bot.
Confirming gold longs with a dollar basket: profit factor 1.04 to 1.18. A filter idea, not a bot.
The TB41 recipe on S&P, Russell and Dow micros: profit factors 1.24, 0.70, 0.93. The edge is Nasdaq-specific.