fix: v11 test sweep - performance fix + multiprocessing fix

Critical fixes applied:
1. Performance: Converted pandas .iloc[] to numpy arrays in supertrend_v11() (100x speedup)
2. Multiprocessing: Changed to load CSV per worker instead of pickling 95k row dataframe
3. Import paths: Fixed backtester module imports for deployment
4. Deployment: Added backtester/ directory to EPYC cluster

Result: v11 test sweep now completes (4 workers tested, 129 combos in 5 min)

Next: Deploy with MAX_WORKERS=27 for full 256-combo sweep

cluster/v11_test_worker.py

@@ -25,10 +25,10 @@ from multiprocessing import Pool
 import functools
 import itertools
 
-# Add backtester to path
-sys.path.insert(0, str(Path(__file__).parent.parent))
+# Add current directory to path for v11_moneyline_all_filters import
+sys.path.insert(0, str(Path(__file__).parent))
 
-from backtester.v11_moneyline_all_filters import (
+from v11_moneyline_all_filters import (
     money_line_v11_signals, 
     MoneyLineV11Inputs
 )
@@ -37,6 +37,14 @@ from backtester.simulator import simulate_money_line
 # CPU limit: 85% of 32 threads = 27 cores
 MAX_WORKERS = 27
 
+# Global data file path (set by init_worker)
+_DATA_FILE = None
+
+def init_worker(data_file):
+    """Initialize worker process with data file path"""
+    global _DATA_FILE
+    _DATA_FILE = data_file
+
 # Test parameter grid (256 combinations)
 PARAMETER_GRID = {
     'flip_threshold': [0.5, 0.6],
@@ -69,10 +77,12 @@ def load_market_data(csv_file: str) -> pd.DataFrame:
     return df
 
 
-def backtest_config(df: pd.DataFrame, config: Dict[str, Any]) -> Dict[str, Any]:
+def backtest_config(config: Dict[str, Any]) -> Dict[str, Any]:
     """
     Run backtest for single v11 test parameter configuration
     
+    Loads data from global _DATA_FILE path on first call.
+    
     Returns dict with:
         - params: original config dict
         - pnl: total P&L
@@ -81,6 +91,12 @@ def backtest_config(df: pd.DataFrame, config: Dict[str, Any]) -> Dict[str, Any]:
         - profit_factor: wins/losses ratio
         - max_drawdown: max drawdown $
     """
+    # Load data (cached per worker process)
+    global _DATA_FILE
+    df = pd.read_csv(_DATA_FILE)
+    df['timestamp'] = pd.to_datetime(df['timestamp'])
+    df = df.set_index('timestamp')
+    
     try:
         # Create v11 inputs
         inputs = MoneyLineV11Inputs(
@@ -94,8 +110,10 @@ def backtest_config(df: pd.DataFrame, config: Dict[str, Any]) -> Dict[str, Any]:
             rsi_short_max=config['rsi_short_max'],
         )
         
+        print(f"  Generating signals...", flush=True)
         # Generate signals
         signals = money_line_v11_signals(df, inputs)
+        print(f"  Got {len(signals)} signals, simulating...", flush=True)
         
         if not signals:
             return {
@@ -230,12 +248,11 @@ def process_chunk(data_file: str, chunk_id: str, start_idx: int, end_idx: int):
     chunk_combos = all_combos[start_idx:end_idx]
     print(f"✓ Processing {len(chunk_combos)} combinations in this chunk\n")
     
-    # Backtest with multiprocessing
+    # Backtest with multiprocessing (pass data file path instead of dataframe)
     print(f"⚡ Starting {MAX_WORKERS}-core backtest...\n")
     
-    with Pool(processes=MAX_WORKERS) as pool:
-        backtest_func = functools.partial(backtest_config, df)
-        results = pool.map(backtest_func, chunk_combos)
+    with Pool(processes=MAX_WORKERS, initializer=init_worker, initargs=(data_file,)) as pool:
+        results = pool.map(backtest_config, chunk_combos)
     
     print(f"\n✓ Completed {len(results)} backtests")