Production Strategies
Live status from Docker logs + state files. All 5 bots share the same GRU ensemble (R069: 5 seeds × 4 folds) — differences are policy, not model.
Live status refreshed at commit time (2026-06-01 08:03 Madrid) · live capital total €107.28
| Bot | Capital | Peak | Trades | State | Regime | DRSI | Cooldown | Last BTC | Last bar | Live |
|---|---|---|---|---|---|---|---|---|---|---|
| V2 | €22.17 | €23.32 | 1 | INVESTED | — | — | — | $73296 | 2026-06-01 06:00:00 | 47d |
| V3 | €23.00 | — | 0 | CASH | — | — | — | $73294 | 2026-06-01 06:00:00 | 47d |
| V4 | €22.56 | €23.57 | 0 | INVESTED | — | — | — | $73294 | 2026-06-01 06:00:00 | 47d |
| V5 | €18.81 | €20.16 | 0 | INVESTED | — | — | — | $73294 | 2026-06-01 06:00:00 | 42d |
| V6 | €20.75 | €20.75 | 0 | INVESTED | — | — | — | $73294 | 2026-06-01 06:00:00 | 42d |
System architecture · GRU 2×128 ensemble (V2-V6 shared)
All 5 production bots share one GRU 2×128 ensemble. The diff between V2 and V6 is only YAML config.
V2 · GRU Ensemble (always-invested)
backtest config: config/production_v2.yaml
Raw 5-seed GRU vote, always-invested. Exits only when ≥3/5 models flag danger. The 'naked model' baseline — useful as a control for V3-V6 attribution: any of them that doesn't beat V2 is not earning its complexity.
WEAKNESS — DON'T COPY IFTime-in-market % — should be ~95%+. Also: min α per fold (undocumented). If min α ≤ B&H on any fold, the whole GRU thesis is fragile.
V3 · GRU + DailyRSI>80 + Trail 10%
backtest config: config/production_v3_hybrid.yaml
GRU + DailyRSI>80 overheat exit + 10% trailing. Designed to surf bulls and step aside on parabolic exhaustion. The astronomical compound number is misleading without fold attribution.
WEAKNESS — DON'T COPY IFSingle-fold contribution share. If >70% of compound comes from one fold, this strategy is one regime away from being worthless. Show the fold decomposition explicitly.
V4 · GRU 3-regime adaptive (V5.4 Robust-5)
backtest config: config/production_v4_hybrid_v5.yaml
Robust-5 adaptive ATR exits keyed to bull/bear/lateral regime. The first strategy in the family with a real per-regime mechanism rather than uniform parameters. R095 confirmed +53% improvement via adaptive exit on this bot.
WEAKNESS — DON'T COPY IFLateral-regime PnL alone. V4's whole bet is regime detection; if its lateral sub-period PnL is < V6's lateral sub-period PnL, the regime mechanism is cosmetic.
V5 · V115_cmp (GRU + peak_drop + ratchet + regime cooldown)
backtest config: config/production_v5_hybrid_v115.yaml
GRU + peak_drop trail + ratchet + per-regime cooldown (greedy bull/bear, 30h punishment lateral). Squeezes maximum compounding from trending phases by re-entering aggressively after pullbacks. Reproduces external agent's V115 result at the dollar.
WEAKNESS — DON'T COPY IFMin α +238% (vs V6's +245%). V5 trades 7pp of consistency for compound. MDD per fold — if any fold MDD > 30%, V5 may not be tradable at scale.
V6 · V66 Cooldown(4,48) — uniform thresholds, extreme cooldown
backtest config: config/production_v6_hybrid_v66.yaml
Same GRU, uniform thresholds, extreme cooldown (4 bars + 48 extra after loss, all regimes). Optimized for the WORST fold, not the best. The benchmark every R&D candidate must beat. Compound +70,576% is the 'price' paid for the +245% min α record.
WEAKNESS — DON'T COPY IFOpportunity cost in mega-bulls. V6 will systematically underperform V5 and B&H in raging bulls. If you want an 'always wins' strategy, V6 is not it.
V7 · V5 V115_cmp + vol-Kelly continuous sizing (margin)
backtest config: config/production_v7_vol_kelly.yaml
Identical V5 cascade for entry/exit. NEW: position size scales with inverse 30-day volatility via Binance USDC borrow. leverage = clip(0.030 / vol_30d_shifted, 0.5, 1.5). Auto-deleverages in high-vol regimes (slip protection). R203 buggy +981K confirmed by R204 honest at 102% retention. R217b confirms V15 aggressive variant (clip[0.5,2.0]) gives +8.8M%/+919 with 0 liquidations — future upgrade candidate after V7 paper trade. V7 inherited 0.0002 BTC + $7.26 USDC as synthetic entry from V3 margin wallet at $73,572.
WEAKNESS — DON'T COPY IFMargin interest cost (~0.45% APR). Liquidation risk in flash crashes — backtest 0 liquidations across 20 stress scenarios but real-world tail untested. Production uses cross margin in V3's sub-account (not V5 — V3 had pre-funded USDC from May-28 test trades).
V8 · LLM Hybrid Trader (V5 cascade + Opus committee filter)
backtest config: PENDING — production_v8_llm_hybrid.yaml (creation blocked by hook)
V5 V115_cmp generates base signal. Every 2h an LLM committee (Opus orchestrator + 3 Sonnet specialists: technical/macro/risk) views multi-timeframe BTC charts (15m/1h/4h/1d with SMA/EMA/BB/RSI) and VETOES entries it finds unsafe. Adds vision-based pattern recognition the GRU lacks.
WEAKNESS — DON'T COPY IFAPI cost ~$3-4/day. Committee can FALSE-NEGATIVE good GRU entries. Latency budget — committee cycle must complete in <90s or block bar processing.
V9 · LLM Pure Trader (no GRU, committee decides directly)
backtest config: PENDING — production_v9_llm_pure.yaml
NO GRU model. Pure LLM committee every 2h drives ENTER/HOLD/EXIT decisions. Position size scales with committee confidence × kelly_fraction_max (0.5). The most experimental of the 3 candidates.
WEAKNESS — DON'T COPY IFVariance — only the LLM filters. No GRU to disagree. Each cycle costs ~$0.30. Slow cadence (every 2h) means missed micro-moves.