📈 SAR Cluster v4 — Final Comparison Report
Generated 2026-06-21 16:53 UTC | 3.5 years × 33 pairs × 41 windows | All strategies ranked
🏆 v4_long is the first profitable version over the full 3.5-year backtest.
Realistic capital: +$1,421 (+1.4%) | Unlimited: +$2,464 (+0.2%)
Key improvement: regime-aware dynamic stoploss (tight 2% in chop, trailing from +5% in trends)
Aggregate Results
Realistic Capital
| Strategy | Total PnL | Avg/Window | Trades | Profitable | Zero |
|---|
| sar_cluster_v1 | -5,316.7 | -129.7 | 2996 | +16/-25 | 0 |
| sar_cluster_v2 | -2,417.6 | -59.0 | 1917 | +20/-21 | 0 |
| sar_cluster_v3 | -1,310.6 | -32.0 | 1528 | +15/-17 | 9 |
| sar_cluster_v4_long | 1,421.2 | 34.7 | 339 | +13/-19 | 9 |
Unlimited Capital
| Strategy | Total PnL | Avg/Window | Trades | Profitable | Zero |
|---|
| sar_cluster_v1 | -4,160.3 | -101.5 | 9128 | +16/-25 | 0 |
| sar_cluster_v2 | -529.1 | -12.9 | 3014 | +20/-21 | 0 |
| sar_cluster_v3 | -373.9 | -9.1 | 2360 | +15/-17 | 9 |
| sar_cluster_v4_long | 2,463.8 | 60.1 | 1026 | +12/-20 | 9 |
v4_long Unlimited — Best & Worst Windows
| Timerange | Trades | PnL | PF |
|---|
| ❌ | 20251017-20251216 | 31 | -729.3 | 0.15 |
| ❌ | 20241221-20250219 | 28 | -724.7 | 0.03 |
| ❌ | 20250917-20251116 | 32 | -624.9 | 0.17 |
| ❌ | 20240425-20240624 | 31 | -612.7 | 0.06 |
| ❌ | 20251216-20260214 | 29 | -584.8 | 0.16 |
| ✅ | 20230928-20231127 | 36 | 1116.4 | 37.54 |
| ✅ | 20240922-20241121 | 34 | 1179.2 | 359.19 |
| ✅ | 20231028-20231227 | 37 | 1263.1 | 88.88 |
| ✅ | 20240126-20240326 | 33 | 1346.8 | 506.87 |
| ✅ | 20241022-20241221 | 54 | 2599.5 | 9.14 |
What's different in v4_long?
- Same entries as v3 — cluster detection + gate + PCA regime filter
- Regime-scoped calculations — no temporal leakage from future data
- Dynamic custom_stoploss replaces fixed trailing stop:
- Strong trend (bull/bear): Allow up to -10% drawdown, trailing activates at +5% profit, locks in 2.5% below peak
- Chop/mixed: Hard floor at -2% loss, trailing at +1.5%, keeps 1% below peak
- Result: Fewer trades (1,026 vs 2,360 in v3), but winners run longer and losers are cut faster