cc56b72df2
CRITICAL FIX (Nov 30, 2025):
- Dashboard showed 'idle' despite 22+ worker processes running
- Root cause: SSH-based worker detection timing out
- Solution: Check database for running chunks FIRST
Changes:
1. app/api/cluster/status/route.ts:
- Query exploration database before SSH detection
- If running chunks exist, mark workers 'active' even if SSH fails
- Override worker status: 'offline' → 'active' when chunks running
- Log: '✅ Cluster status: ACTIVE (database shows running chunks)'
- Database is source of truth, SSH only for supplementary metrics
2. app/cluster/page.tsx:
- Stop button ALREADY EXISTS (conditionally shown)
- Shows Start when status='idle', Stop when status='active'
- No code changes needed - fixed by status detection
Result:
- Dashboard now shows 'ACTIVE' with 2 workers (correct)
- Workers show 'active' status (was 'offline')
- Stop button automatically visible when cluster active
- System resilient to SSH timeouts/network issues
Verified:
- Container restarted: Nov 30 21:18 UTC
- API tested: Returns status='active', activeWorkers=2
- Logs confirm: Database-first logic working
- Workers confirmed running: 22+ processes on worker1, workers on worker2
105 lines
3.9 KiB
Markdown
105 lines
3.9 KiB
Markdown
# V9 Diagnostic Results Summary
|
|
**Date:** November 29, 2025
|
|
**Data:** 95,617 5-minute bars (SOLUSDT, Jan-Nov 2024)
|
|
|
|
## 🚨 CRITICAL FINDINGS
|
|
|
|
### 1. **BASELINE IS LOSING MONEY**
|
|
- **Baseline PnL:** -$1,532.30 (1,663 trades)
|
|
- **Win Rate:** 0.6% (essentially all losses!)
|
|
- **This explains the "parameter insensitivity"** - when strategy loses on EVERY trade, parameters don't matter much
|
|
|
|
### 2. **momentum_min_adx Parameter BROKEN**
|
|
```
|
|
momentum_min_adx=18.0: 1663 trades, $-1532.30 PnL
|
|
momentum_min_adx=21.0: 1663 trades, $-1532.30 PnL ← IDENTICAL
|
|
momentum_min_adx=24.0: 1663 trades, $-1532.30 PnL ← IDENTICAL
|
|
momentum_min_adx=27.0: 1663 trades, $-1532.30 PnL ← IDENTICAL
|
|
```
|
|
**Status:** 🔴 NO EFFECT - Parameter is NOT being applied or is overridden
|
|
|
|
### 3. **Other Parameters Show Minimal Effect**
|
|
- **flip_threshold:** 1662-1663 trades (0.1% variation), PnL: -$1,185 to -$1,532
|
|
- **cooldown_bars:** 1660-1664 trades (0.2% variation), PnL: -$1,408 to -$1,859
|
|
- **ma_gap_threshold:** 1662-1663 trades (0.1% variation), PnL: -$1,185 to -$1,532
|
|
|
|
**Signal counts barely change** - most parameters have almost zero effect on trade generation.
|
|
|
|
## 📊 Comparison to Sweep Results
|
|
|
|
**Exhaustive Sweep (EPYC):**
|
|
- Best Result: $498.12 PnL, 568 trades, 61.09% WR
|
|
- Configuration: Different from baseline
|
|
|
|
**Diagnostic Test (Local):**
|
|
- Baseline: -$1,532.30 PnL, 1,663 trades, 0.6% WR
|
|
- Best: -$1,514.75 PnL, 1,663 trades, 0.6% WR
|
|
|
|
## 🤔 Why The Discrepancy?
|
|
|
|
### Hypothesis 1: Data Mismatch
|
|
- **EPYC used:** Aug 1 - Nov 28, 2024 (34,273 candles - mentioned in DUAL_SWEEP_README.md)
|
|
- **Local used:** Jan 1 - Nov 28, 2024 (95,617 candles - full year)
|
|
- **Impact:** Different time periods = different market conditions = different results
|
|
|
|
### Hypothesis 2: Configuration Mismatch
|
|
- EPYC sweep might be using different TradeConfig settings
|
|
- Position size, max bars per trade, or other simulator settings might differ
|
|
|
|
### Hypothesis 3: Strategy Implementation Difference
|
|
- Backtester `simulate_money_line()` might not match live v9 indicator
|
|
- Parameters might not map correctly between TradingView and Python
|
|
|
|
## 🎯 Action Items
|
|
|
|
### IMMEDIATE (Before Any Optimization):
|
|
|
|
1. **✅ VERIFY DATA ALIGNMENT**
|
|
```bash
|
|
# Download exact same date range as EPYC
|
|
python3 scripts/export_binance_ohlcv.py \
|
|
--symbol SOLUSDT --interval 5m \
|
|
--start 2024-08-01 --end 2024-11-28 \
|
|
--output backtester/data/solusdt_5m_aug_nov.csv
|
|
|
|
# Re-run diagnostics on matched dataset
|
|
./run_comprehensive_diagnostics.sh backtester/data/solusdt_5m_aug_nov.csv
|
|
```
|
|
|
|
2. **VERIFY SIMULATOR SETTINGS**
|
|
- Check if EPYC sweep uses different position_size or max_bars_per_trade
|
|
- Compare TradeConfig between sweep script and diagnostic scripts
|
|
|
|
3. **FIX momentum_min_adx BUG**
|
|
- Investigate money_line_signals() to find why ADX parameter is ignored
|
|
- This is likely why all sweep configs produced similar results
|
|
|
|
4. **FIX EXTREME BUGS**
|
|
- Fix load_csv() call in test_extreme_configs() (missing symbol/timeframe)
|
|
- Fix SimulatedTrade.pnl attribute access in trade_analysis.py
|
|
|
|
### AFTER VERIFICATION:
|
|
|
|
5. **If Data Mismatch Confirmed:**
|
|
- Use Aug-Nov 2024 dataset for all future analysis
|
|
- Understand why Q1-Q3 2024 was so terrible (bear market?)
|
|
|
|
6. **If Simulator Bug Confirmed:**
|
|
- Fix Python backtester to match TradingView v9 exactly
|
|
- Validate against known live trades
|
|
|
|
7. **Parameter Optimization:**
|
|
- Only optimize AFTER baseline is profitable on test data
|
|
- No point optimizing if strategy loses money fundamentally
|
|
|
|
## 💡 Key Insight
|
|
|
|
**You can't optimize a fundamentally losing strategy.**
|
|
|
|
If v9 baseline loses $1,532 on full-year data but makes $498 on Aug-Nov subset, either:
|
|
- A) Aug-Nov was a favorable period (cherry-picked results)
|
|
- B) Jan-Jul market was unfavorable for momentum strategies (bear market)
|
|
- C) Backtester doesn't match production v9 indicator
|
|
|
|
**Must resolve this before any parameter tuning!**
|