DECODED STRATEGY — Beat (January 2026)
6 ML models + 5 rule systems | +3,015 SOL (345% of Beat pump.fun PnL) | 47.9% WR | 45,335 trades
Scorecard vs Beat
| metric | score | grade | bar | detail |
|---|
| Hold duration | 96.0% | A | | Median diff -2 frames vs Beat |
| Entry (token selection) | 67.7% | B | | Catch 68% of Beat's entries |
| Re-entry overlap | 65.1% | B | | 65% of Beat's re-entered tokens |
| Outcome match | 51.8% | C | | 52% same win/loss on shared tokens |
| Entry timing | 40.0% | C | | Median -12 frames vs Beat |
| Sizing (exact) | 23.8% | D | | 24% within 0.05 SOL |
| Exit timing (±5f) | 8.9% | F | | Only 9% within ±5 frames (23.9% with autoresearch) |
ML Models
| model | type | metric | grade | features | top features | notes |
|---|
| Filter | XGBoost | AUC 1.000 | B | 29 | momentum_score (32%), tpm_10 (30%), buy_ratio_5 (12%), price_chg_5 (10%) | Verified on 610K random tokens — legit, not artifact. Rejects 99.9% of pump.fun garbage. |
| Entry | XGBoost | AUC 0.826 | B | 27 | sell_streak (36%), buy_streak, unique_traders, time_since_prev, holder_count | Catches 68% of Beat's entries. Enters 12 frames before Beat (median). +2,163 SOL early-entry advantage. |
| Sizing (v2) | XGBoost 9-class | 81% train acc | D | 70 | open_positions (9.7%), sol_deployed (6.9%), buy_size_ratio_prev (5.8%) | 9 discrete sizes matching Beat's grid. 24% exact match live. Sizing is portfolio-driven. |
| Runner | XGBoost | AUC 0.799 | — | 58 | sol_to_migration (6.5%), creator_rebuy_count, trades_in_slot, buy_count, range_pct_10 | Feeds into exit model as feature. Gates re-entries (runner >= 0.40 required). |
| Exit | XGBoost + rules | AUC 0.860 | F | 64 | unrealized_pnl_ratio (7.2%), tpm_10 (5.5%), time_since_prev, runner_score (1.5%) | 8.9% within ±5f of Beat (improved to 23.9% via autoresearch asymmetric labels). Threshold: 0.45. |
| Post-mig Entry | XGBoost | AUC 0.602 | — | 30 | Pump.fun features (weak — can't predict Raydium behavior) | Not used in production. Pump.fun features don't predict Raydium. |
Rule Systems
| system | when | logic | notes |
|---|
| Regime | Hourly | HOT: grad_rate >= 0.55% & vol > 35K | COLD: grad_rate < 0.30% | WARM: else | NOT used in simulation (tested, hurt PnL). 22/31 days are COLD. Fixed WARM params outperform. |
| Re-entry (6-9) | After sell_signal | Recovery 2% + runner >= 0.40 + max 3 buys. Buy #2 = 4x, #3 = 6x | +1,453 SOL from 24,900 re-entries. Buy #2 weakest at 42.7% WR. |
| Staleness | Every frame | time_since_prev >= 30s AND frames_since_entry >= 3 → FORCE SELL | -617 SOL, 20% WR. Wrong 46% of the time (tokens recover). Biggest WR drag. |
| Graduation hold | While holding | curve >= 80% AND velocity > 0 AND sell_score < 0.5 → HOLD | Overrides sell signal near migration. 25 trades, 92% WR. |
| Timeout | While holding | frames_since_entry >= 200 → FORCE SELL or write off | 867 force + 406 dead. Beat's P99 hold = 163 frames. |
Pipeline Flow
Geyser Stream (BlockRazor gRPC) → Accumulator Engine (O(1)/trade)
→ Filter (AUC 1.0, frame 8) — rejects 99.9% of pump.fun
→ Entry (AUC 0.826, every frame) — threshold 0.50
→ Sizing (9-class v2: 1x/2x/2.5x/3x/4x/5x/6x/9x/12.5x)
→ Hold/Monitor:
Runner (AUC 0.799) feeds into Exit (AUC 0.860)
Priority: staleness 30s → timeout 200f → graduation hold → sell_score >= 0.45
→ Re-entry chain (fixed WARM: 2% recovery, runner >= 0.40, max 3 buys)
→ Execution via 0slot (anti-MEV, ~0.001 SOL tip/trade)
Pipeline speed: 0.6ms compute per token (1,651 tokens/sec)Audit Findings
1. Filter is legit — AUC 1.0 on 610K unseen tokens, not an artifact
2. +2,163 SOL (72%) of first-buy PnL from entering before Beat pushes price up
3. If we entered at Beat's price instead of earlier: -601 SOL (net loss on first buys)
4. Entry model fires on 34% of random tokens if filter bypassed (filter catches this)
5. Winners vs Losers at frame 8: AUC 0.56 (can't tell them apart early)
6. Stale exits: 20% WR, wrong 46% of the time — tokens recover after we panic sell
7. Non-stale trades already at 50.0% WR — stale is the only WR drag
8. Exit model improved from 8.9% → 23.9% via autoresearch (asymmetric labels)
9. Regime-adjusted re-entry HURTS PnL — fixed WARM outperforms
10. max_buys=5 adds +510 SOL over max_buys=3
Vulnerabilities
CRITICAL:
• Slippage sensitivity: +1% worse slippage flips 2,058 thin wins to losses (WR -4.5%)
• Early entry dependency: 87% of first-buy PnL from entries before frame 20
• Model trained on Jan 2026 — may not generalize to current market
MODERATE:
• Stale exits lose -617 SOL (46% of force-sells are wrong)
• Dead tokens: -313 SOL from 406 total losses
• Network latency: 50-200ms from decision to execution
MITIGATED:
• Kill switch + daily loss limit + max drawdown (P0 fail-safes)
• Paper mode default — must explicitly enable live trading
• Wallet separation recommended
Data Integrity
• Beat's trades excluded from all training (accumulators, rolling windows, geyser)
• Source parquet NEVER modified — all cleaning via SQL joins
• Temporal split: Jan 1-24 train, Jan 25-31 test
• Real slippage from Beat's trades: Probe 4.78%, Medium 2.52%, Large 1.86%