From 7e8722aa57c21c89be45d63712e748a1f6bdaacb Mon Sep 17 00:00:00 2001
From: Dobromir Popov <dobromir.popov@gmail.com>
Date: Mon, 8 Dec 2025 19:35:58 +0200
Subject: [PATCH] no data rail guide

---
 ANNOTATE/core/real_training_adapter.py | 37 +++++++++++++++++++++++---
 1 file changed, 33 insertions(+), 4 deletions(-)

diff --git a/ANNOTATE/core/real_training_adapter.py b/ANNOTATE/core/real_training_adapter.py
index e08bf59..a98a828 100644
--- a/ANNOTATE/core/real_training_adapter.py
+++ b/ANNOTATE/core/real_training_adapter.py
@@ -930,6 +930,17 @@ class RealTrainingAdapter:
                 # Create market state snapshot at this candle
                 hold_market_state = self._create_market_state_snapshot(market_state, idx)
                 
+                # Validate that the snapshot has enough data
+                has_enough_data = True
+                for tf, tf_data in hold_market_state.get('timeframes', {}).items():
+                    if len(tf_data.get('close', [])) < 50:  # Minimum required
+                        has_enough_data = False
+                        break
+                
+                if not has_enough_data:
+                    logger.debug(f"      Skipping HOLD sample at idx {idx} - insufficient data")
+                    continue
+                
                 # Calculate current unrealized PnL at this point
                 entry_price = expected_outcome.get('entry_price', 0)
                 current_price = timeframes['1m']['close'][idx] if idx < len(timeframes['1m']['close']) else entry_price
@@ -1043,18 +1054,22 @@ class RealTrainingAdapter:
                 logger.debug(f"      Estimated exit index: {exit_index} ({candles_in_trade} candles)")
             
             # Create NO_TRADE samples: 5 before entry + 5 after exit
+            # CRITICAL: Only create samples that have enough historical data (600 candles)
+            required_candles = 600
             negative_indices = []
             
-            # 5 candles BEFORE entry
+            # 5 candles BEFORE entry (but only if we have enough history)
             for offset in range(1, samples_before + 1):
                 idx = entry_index - offset
-                if 0 <= idx < len(timestamps):
+                # Check if we have enough candles before this index
+                if idx >= required_candles - 1 and idx < len(timestamps):
                     negative_indices.append(('before_entry', idx))
             
-            # 5 candles AFTER exit
+            # 5 candles AFTER exit (should have enough history since exit is after entry)
             for offset in range(1, samples_after + 1):
                 idx = exit_index + offset
-                if 0 <= idx < len(timestamps):
+                # Check if we have enough candles before this index
+                if idx >= required_candles - 1 and idx < len(timestamps):
                     negative_indices.append(('after_exit', idx))
             
             # Create negative samples
@@ -1062,6 +1077,17 @@ class RealTrainingAdapter:
                 # Create a market state snapshot at this timestamp
                 negative_market_state = self._create_market_state_snapshot(market_state, idx)
                 
+                # Validate that the snapshot has enough data
+                has_enough_data = True
+                for tf, tf_data in negative_market_state.get('timeframes', {}).items():
+                    if len(tf_data.get('close', [])) < 50:  # Minimum required
+                        has_enough_data = False
+                        break
+                
+                if not has_enough_data:
+                    logger.debug(f"      Skipping negative sample at idx {idx} - insufficient data")
+                    continue
+                
                 negative_sample = {
                     'market_state': negative_market_state,
                     'action': 'HOLD',  # No action
@@ -2492,6 +2518,9 @@ class RealTrainingAdapter:
                 
                 OPTIMIZATION: Batches are already on GPU and grouped for efficient processing.
                 Each mini-batch contains 5 samples for better GPU utilization.
+                
+                IMPORTANT: Yields the same batch objects across epochs (no copying).
+                The train_step method should not modify batch contents in-place.
                 """
                 for batch in grouped_batches:
                     yield batch