TransLSTM-Predictor: State-of-the-art Stock Prediction System

Overview

TransLSTM-Predictor is a quantitative trading model that combines CNN, Bi-LSTM, and Transformer architectures to achieve state-of-the-art (SOTA) accuracy in stock movement forecasting.

Unlike traditional price predictors, this system focuses on Percentage Returns Prediction, utilizing advanced ensemble methods and rigorous validation strategies to provide reliable trading signals.

Advanced Features

• Return-Based Prediction (SOTA Strategy): Predicts % daily returns instead of absolute prices, significantly improving model stability and generalizability across different price scales.
• CNN-LSTM-Transformer Hybrid:
  • CNN: Extracts local spatial features (price patterns).
  • Bi-LSTM: Captures long-term temporal dependencies.
  • Transformer: Multi-head attention mechanism for complex global relationships.
• Quantitative Validation Suite:
  • Walk-forward Validation: Multi-fold time-series cross-validation to prevent overfitting to specific market regimes.
  • Ensemble Learning: Averages predictions from multiple independently trained models to reduce variance and improve robustness.
• Feature Engineering: Includes high-impact technical indicators:
  • Trend: MA(7, 21), MACD.
  • Volatility: Bollinger Bands, ATR (Average True Range).
  • Volume/Momentum: RSI, OBV (On-Balance Volume).
• Financial Backtesting: Integrated simulator to evaluate the economic performance of the model (Total Return, Sharpe Ratio, MDD, Win Rate).
• Dual Scaler System: Separate normalization logic for features and targets to eliminate data leakage and price explosion issues during reconstruction.

Requirements

pip install -r requirements.txt

How to Use

1. Place your data: Put your stock data CSV files in the data/ directory (Columns: date, open, high, low, close, volume).
2. Run the pipeline:

   python main.py data/YOUR_STOCK_DATA.csv

The system will orchestrate:

1. Data loading & Feature Extraction
2. Walk-forward Validation (Multi-fold training)
3. Ensemble Prediction
4. Backtesting analysis
5. 30-day Future Forecasting & Plotting

Configuration (config/config.py)

• PREDICT_RETURNS: Toggle between price/return prediction modes.
• ENSEMBLE_SIZE: Number of parallel models to train.
• WALK_FORWARD_FOLDS: Number of folds for rigorous validation.
• Model Hyperparameters: Adjust Transformer heads, layers, and LSTM units.

Output

• Model Files: ./results/models/
• Visualization Plots: ./results/plots/ (Includes Prediction vs Actual & Backtest Equity Curve)
• Predictions: ./results/predictions/