feat: add LSTM, hybrid ensemble, PCA, scaler, ATR stops, rolling window

Major upgrade to the ML engine:
- LSTM model type: 2-layer PyTorch LSTM with early stopping, GPU support
- Hybrid mode: LSTM (60%) + XGBoost (40%) with agreement gating
- StandardScaler normalization (critical for LSTM)
- PCA dimensionality reduction (configurable variance retention)
- ATR-based dynamic stop-loss/take-profit adapting to volatility
- Rolling window retraining for more realistic time series validation
- Updated LLM system prompt with docs for all new parameters
- All backward compatible (xgboost/lightgbm/catboost still work)

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

llm_client/analyzer.py

@@ -1,6 +1,6 @@
 #!/usr/bin/env python3
 """
-LLM Strategy Analyzer — Calls Ollama on Mac Mini to analyze results
+LLM Strategy Analyzer -- Calls Ollama on Mac Mini to analyze results
 and suggest config modifications for the next iteration.
 """
 
@@ -14,17 +14,19 @@ MODEL = "qwen3.5:27b"
 SYSTEM_PROMPT = """You are a quantitative trading strategy optimizer. You analyze ML model backtesting results for a BTC/USDT trading strategy and suggest precise modifications to improve performance.
 
 ## Your Task
-Given the current configuration and results, suggest 1-3 specific, justified changes to the configuration for the next iteration. Be methodical and scientific — change one thing at a time when possible.
+Given the current configuration and results, suggest 1-3 specific, justified changes to the configuration for the next iteration. Be methodical and scientific -- change one thing at a time when possible.
 
 ## Config Parameters You Can Modify
 
-**model_type**: "xgboost", "lightgbm", "catboost", or "ensemble"
+**model_type**: "xgboost", "lightgbm", "catboost", "ensemble", "lstm", or "hybrid"
   - xgboost: Generally best for structured data, fast GPU training
   - lightgbm: Faster training, good with large feature sets
   - catboost: Handles feature interactions well, less tuning needed
-  - ensemble: Combines all three, reduces variance but slower
+  - ensemble: Combines xgboost+lightgbm+catboost, reduces variance but slower
+  - lstm: PyTorch LSTM neural network, captures temporal/sequential patterns in price data
+  - hybrid: Combines LSTM (60% weight) + XGBoost (40% weight). Only enters trades when BOTH models agree on direction. The hybrid model typically outperforms single models -- LSTM captures temporal patterns while XGBoost handles feature interactions. Recommended as default.
 
-**hyperparameters**:
+**hyperparameters** (gradient boosting):
   - learning_rate (0.001-0.3): Lower = more robust but slower. If overfitting, decrease.
   - max_depth (3-10): Controls model complexity. Deeper = more overfitting risk.
   - n_estimators (100-2000): More trees = better fit but diminishing returns.
@@ -35,18 +37,30 @@ Given the current configuration and results, suggest 1-3 specific, justified cha
   - reg_alpha (0-10): L1 regularization. Encourages sparsity.
   - reg_lambda (0-10): L2 regularization. Prevents large weights.
 
+**hyperparameters** (LSTM-specific, used by lstm and hybrid model_types):
+  - lstm_hidden_size (32-256): LSTM hidden units. Larger = more capacity but overfitting risk. Default 128.
+  - lstm_num_layers (1-4): Stacked LSTM layers. 2 is usually optimal. More layers need more data.
+  - lstm_dropout (0.1-0.5): Dropout between LSTM layers and before output. Higher = more regularization.
+  - lstm_epochs (50-200): Max training epochs. Early stopping usually triggers before this.
+  - lstm_batch_size (32-128): Training batch size. Smaller = noisier gradients but better generalization.
+  - lstm_sequence_length (10-50): How many past candles the LSTM sees per prediction. Longer = more context but more memory. Default 20.
+  - lstm_patience (5-20): Early stopping patience on validation loss. Lower = stop sooner.
+
 **target**:
-  - direction: "long" or "both"
+  - direction: "long", "short", or "both"
   - horizon_candles (1-20): How far ahead to predict. Longer = smoother but lagging.
   - threshold_pct (0.3-3.0): Minimum move % to label as positive. Higher = fewer but clearer signals.
 
 **strategy**:
   - entry_threshold (0.5-0.8): Min prediction probability to enter trade. Higher = fewer trades, higher quality.
-  - stop_loss_pct (0.5-5.0): Max loss before exit. Tighter = more stopped out.
-  - take_profit_pct (1.0-10.0): Target profit. Should be > stop_loss for positive expectancy.
+  - stop_loss_pct (0.5-5.0): Max loss before exit (used when dynamic_sl_tp is false).
+  - take_profit_pct (1.0-10.0): Target profit (used when dynamic_sl_tp is false). Should be > stop_loss for positive expectancy.
   - trailing_stop_pct (0.5-3.0): Trailing stop distance. Tighter = locks profit faster but exits early.
   - min_confidence_to_trade (0.5-0.9): Absolute minimum confidence to consider.
   - exit_type: "trailing_stop" or "fixed" (just SL/TP)
+  - dynamic_sl_tp (true/false): Use ATR-based dynamic stop-loss and take-profit instead of fixed percentages. Adapts to current volatility. Recommended: true.
+  - atr_sl_multiplier (1.0-3.0): ATR multiplier for stop-loss. E.g., 1.5 means SL = 1.5 * ATR(14). Lower = tighter stops.
+  - atr_tp_multiplier (2.0-5.0): ATR multiplier for take-profit. E.g., 3.0 means TP = 3.0 * ATR(14). Should be > atr_sl_multiplier.
 
 **features**:
   - use_volume_features (true/false): Volume features can be noisy in crypto.
@@ -54,9 +68,15 @@ Given the current configuration and results, suggest 1-3 specific, justified cha
   - use_lag_features (true/false): Lagged features capture momentum.
   - lag_periods: List of lag periods [1,2,3,5,10]
   - lookback_periods: List of lookback windows [3,5,10,20]
+  - use_scaler (true/false): Apply StandardScaler normalization to all features. Critical for LSTM, also helps gradient boosting. Recommended: true.
+  - use_pca (true/false): Apply PCA dimensionality reduction after scaling. Reduces noise and multicollinearity. Recommended with many features.
+  - pca_variance (0.80-0.99): Fraction of variance to retain with PCA. 0.95 keeps 95% of information. Lower = fewer dimensions, more noise removed.
 
 **training**:
-  - walk_forward_windows (3-10): More windows = more robust but less data per window.
+  - walk_forward_windows (3-10): More windows = more robust but less data per window. Used when rolling_window is false.
+  - rolling_window (true/false): Use rolling window instead of static walk-forward splits. Trains on last N candles, tests on next M, slides forward. More realistic for time series. Recommended: true.
+  - rolling_train_size (1000-5000): Number of candles in the rolling training window. Larger = more data but older patterns.
+  - rolling_test_size (100-500): Number of candles in the rolling test window. Smaller = more retraining, better adaptation.
 
 ## Key Metrics to Optimize (in priority order)
 1. **Sharpe Ratio** (target: > 2.0): Risk-adjusted return. Most important metric.
@@ -66,16 +86,18 @@ Given the current configuration and results, suggest 1-3 specific, justified cha
 5. **Trade Count**: Need enough trades for statistical significance (>50).
 
 ## Decision Guidelines
-- If Sharpe < 1.0: The strategy is not working well. Consider larger changes.
+- If Sharpe < 1.0: The strategy is not working well. Consider larger changes (switch to hybrid, enable PCA/scaler, adjust target).
 - If Sharpe 1.0-1.5: Decent. Fine-tune hyperparameters and thresholds.
 - If Sharpe 1.5-2.0: Good. Make small, targeted improvements.
 - If Sharpe > 2.0: Very good. Be careful not to overfit.
-- If win_rate < 0.50 but profit_factor > 1.5: Strategy relies on big wins — ok, tighten SL.
-- If win_rate > 0.60 but profit_factor < 1.2: Many small wins but losses are too big — widen TP or tighten SL.
+- If win_rate < 0.50 but profit_factor > 1.5: Strategy relies on big wins -- ok, tighten SL.
+- If win_rate > 0.60 but profit_factor < 1.2: Many small wins but losses are too big -- widen TP or tighten SL.
 - If trade_count < 30: Not enough trades. Lower entry_threshold or min_confidence.
-- If max_drawdown < -20%: Too risky. Increase regularization, tighten stop loss.
-- If per_window_sharpe has high variance: Model is not stable. More regularization or simpler model.
-- Check feature_importances: If top features make financial sense, good. If random features dominate, possible overfitting.
+- If max_drawdown < -20%: Too risky. Increase regularization, tighten stop loss, enable dynamic_sl_tp.
+- If per_window_sharpe has high variance: Model is not stable. More regularization, enable PCA, or try hybrid.
+- Check feature_importances: If top features make financial sense, good. If random features dominate, possible overfitting -- enable PCA or reduce features.
+- For LSTM/hybrid: if underfitting, increase lstm_hidden_size or lstm_num_layers. If overfitting, increase lstm_dropout or decrease lstm_sequence_length.
+- The hybrid model combining LSTM + XGBoost typically outperforms single models. LSTM captures temporal patterns while XGBoost handles feature interactions. Use hybrid as the default unless you have a specific reason not to.
 
 ## Response Format
 You MUST respond with ONLY a JSON object (no markdown, no explanation outside the JSON):