From 907a7d6224cf479f7df319cd0ee2056a7e74beef Mon Sep 17 00:00:00 2001
From: Dobromir Popov <dobromir.popov@gateway.one>
Date: Tue, 4 Nov 2025 13:30:39 +0200
Subject: [PATCH] adding current position PnL to the T model

---
 NN/models/advanced_transformer_trading.py | 32 +++++++++++++++--------
 1 file changed, 21 insertions(+), 11 deletions(-)

diff --git a/NN/models/advanced_transformer_trading.py b/NN/models/advanced_transformer_trading.py
index ca88e94..96e0439 100644
--- a/NN/models/advanced_transformer_trading.py
+++ b/NN/models/advanced_transformer_trading.py
@@ -355,6 +355,18 @@ class AdvancedTradingTransformer(nn.Module):
         self.tech_projection = nn.Linear(config.tech_features, config.d_model)
         self.market_projection = nn.Linear(config.market_features, config.d_model)
         
+        # Position state projection - properly learns to embed position info
+        # Input: [has_position, pnl, size, entry_price_norm, time_in_position] = 5 features
+        self.position_projection = nn.Sequential(
+            nn.Linear(5, config.d_model // 4),  # 5 -> 256
+            nn.GELU(),
+            nn.Dropout(config.dropout),
+            nn.Linear(config.d_model // 4, config.d_model // 2),  # 256 -> 512
+            nn.GELU(),
+            nn.Dropout(config.dropout),
+            nn.Linear(config.d_model // 2, config.d_model)  # 512 -> 1024
+        )
+        
         # Positional encoding
         if config.use_relative_position:
             self.pos_encoding = RelativePositionalEncoding(config.d_model)
@@ -516,19 +528,17 @@ class AdvancedTradingTransformer(nn.Module):
         
         # Add position state if provided - critical for loss minimization and profit taking
         if position_state is not None:
-            # Project position state to model dimension and add to all sequence positions
-            # This allows the model to condition all predictions on current position state
-            position_emb = torch.tanh(position_state)  # Normalize to [-1, 1]
-            position_emb = position_emb.unsqueeze(1).expand(batch_size, seq_len, -1)  # (batch, seq_len, 5)
+            # Project position state through learned embedding network
+            # Input: [batch, 5] -> Output: [batch, d_model]
+            position_emb = self.position_projection(position_state)  # [batch, d_model]
             
-            # Pad to match model dimension if needed
-            if position_emb.size(-1) < self.config.d_model:
-                padding = torch.zeros(batch_size, seq_len, self.config.d_model - position_emb.size(-1), 
-                                     device=position_emb.device, dtype=position_emb.dtype)
-                position_emb = torch.cat([position_emb, padding], dim=-1)
+            # Expand to sequence length and add as bias to all positions
+            # This conditions the entire sequence on current position state
+            position_emb = position_emb.unsqueeze(1).expand(batch_size, seq_len, -1)  # [batch, seq_len, d_model]
             
-            # Add position state as a bias to the embeddings
-            x = x + position_emb[:, :, :self.config.d_model]
+            # Add position embedding to the combined embeddings
+            # This allows the model to modulate its predictions based on position state
+            x = x + position_emb
         
         # Add positional encoding
         if isinstance(self.pos_encoding, RelativePositionalEncoding):