| 每輪取前幾名 | 分數上限 | 目標%(權益) | 移動停利%(權益) | 單筆等效複利率 | 軌道獲利率 | ROI | 停損觸發率 |
|---|---|---|---|---|---|---|---|
| 1 | 不設上限 | 0.050% | 5.868% | -0.676% | 29.2% | +0.12% | 1.1% |
| 1 | 不設上限 | 0.065% | 6.271% | -0.684% | 29.2% | +0.13% | 1.1% |
| 1 | 不設上限 | 0.067% | 5.919% | -0.685% | 29.2% | +0.11% | 1.1% |
| 1 | 不設上限 | 0.095% | 5.940% | -0.685% | 29.2% | +0.11% | 1.1% |
| 1 | 不設上限 | 0.065% | 5.950% | -0.687% | 29.2% | +0.11% | 1.1% |
| 1 | 不設上限 | 0.050% | 5.957% | -0.689% | 29.2% | +0.11% | 1.1% |
| 1 | 不設上限 | 0.062% | 5.139% | -0.689% | 29.2% | +0.09% | 1.1% |
| 1 | 不設上限 | 0.065% | 5.174% | -0.690% | 29.2% | +0.09% | 1.1% |
| 1 | 不設上限 | 0.079% | 5.967% | -0.691% | 29.2% | +0.11% | 1.1% |
| 1 | 不設上限 | 0.068% | 5.696% | -0.692% | 29.2% | +0.10% | 1.1% |
| 1 | 不設上限 | 0.073% | 5.505% | -0.694% | 29.2% | +0.09% | 1.1% |
| 1 | 不設上限 | 0.060% | 5.223% | -0.694% | 29.2% | +0.09% | 1.1% |
| 1 | 不設上限 | 0.060% | 5.511% | -0.695% | 29.2% | +0.09% | 1.1% |
| 1 | 不設上限 | 0.055% | 5.230% | -0.695% | 29.2% | +0.09% | 1.1% |
| 1 | 不設上限 | 0.064% | 5.983% | -0.696% | 29.2% | +0.10% | 1.1% |
| 1 | 不設上限 | 0.064% | 5.034% | -0.696% | 29.2% | +0.09% | 1.1% |
| 1 | 不設上限 | 0.058% | 5.118% | -0.696% | 29.2% | +0.09% | 1.1% |
| 1 | 不設上限 | 0.060% | 5.041% | -0.696% | 29.2% | +0.09% | 1.1% |
| 1 | 不設上限 | 0.062% | 5.128% | -0.697% | 29.2% | +0.08% | 1.1% |
| 1 | 不設上限 | 0.058% | 5.087% | -0.697% | 29.2% | +0.08% | 1.1% |
綠色標示為最佳試驗。這份報告用貝氏優化(Optuna函式庫的TPE演算法)對「每輪取前幾名、分數 上限、目標%、移動停利%」這4個維度一起搜尋，而不是像舊版報告那樣依序固定其它維度只掃一個 維度——2026-09-07實測發現這幾個維度彼此有交互作用：例如目標%/移動停利%從10%/2%改成 更低的值之後，原本在10%/2%底下最好的分數上限反而變差、改成不設上限才是新的最佳，證明 「各自最佳」不等於「聯合最佳」，才改用這個方法。因為訓練/排名只需要做一次、每次試驗只是 輕量地切候選+模擬出場，通常一兩百次試驗就能收斂到穩定的聯合最佳解，比窮舉網格快很多。 若同時有設定「移動停利最小回調率限制」，代表搜尋出來的trail_pct已經確保換算成標的價格 層級的回調率不會比交易所跟蹤委託單能設定的最小檔位還細，這組結果應該可以直接對應到真實 下單用的跟蹤委託單參數。僅供參考，非投資建議。