LSTM for Price Prediction

This repository implements a sequence-to-one LSTM model to predict the next-day Close price of a stock using daily OHLCV data enriched with technical indicators (RSI, MACD, Bollinger Bands). It includes a reproducible training pipeline, Bayesian hyperparameter tuning, evaluation/plotting, and single-step inference.

---

What’s Inside

• Pure LSTM focus (no mixed baselines) with a tidy Keras/TensorFlow pipeline.
• Feature engineering with ta library (RSI, MACD, Bollinger Bands).
• Windowed sequences (configurable lookback) → predict next-day Close.
• Bayesian Optimization for LSTM hyperparameters (skopt).
• Saved artifacts for reuse: results/lstm_model.h5, results/scaler.pkl, and tuning results.

---

Repository structure

├─ data/
│ └─ stock.csv # Daily OHLCV (Date, Symbol, Open, High, Low, Close, …)
├─ results/
│ ├─ bayes_opt_result.pkl # skopt result object (after tuning)
│ ├─ lstm_model.h5 # trained Keras model
│ └─ scaler.pkl # fitted MinMaxScaler for features
├─ src/
│ ├─ features.py # add_technical_indicators, select_features
│ └─ model.py # build_lstm_model(...)
├─ backtest.py # (optional) placeholder for strategy-level analysis
├─ bayes_tuning.py # Bayesian hyperparameter search (skopt)
├─ config.yaml # training & model hyperparameters
├─ evaluate.py # loads model, plots Actual vs Predicted, RMSE
├─ predict.py # single-step next-day Close prediction
├─ requirements.txt
└─ train.py

Data & Features

• Input: data/stock.csv (includes Date, OHLC, Volume, etc.).
• Indicators computed in src/features.py:
  • RSI, MACD (diff), BB_High, BB_Low.
• Final model features (configurable):
  Open, High, Low, Close, Volume, RSI, MACD, BB_High, BB_Low.

Configuration (config.yaml)

Key fields:

• features: ordered list used for scaling + inverse transform.

• window_size: lookback length (days).

• lstm_units, dropout_rate, dense_units

• batch_size, epochs, validation_split

• optimizer, loss

Train

python train.py

What it does:

1. Loads data/stock.csv.

2. Adds indicators → selects features → drops NaN.

3. Scales with MinMaxScaler (fit on all configured features).

4. Builds sequences of length window_size to predict next-day Close.

5. Trains the LSTM with early stopping and saves:

   • results/lstm_model.h5

   • results/scaler.pkl

Evaluate (RMSE + Plot)

python evaluate.py

• Loads model + scaler from results/.

• Rebuilds sequences (fresh split).

• Prints Test RMSE and shows Actual vs Predicted Close plot.

Predict Next-Day Close

python predict.py

• Uses the latest window_size days from data/stock.csv.

• Prints: Predicted Next Day Close Price: <value>

Hyperparameter Tuning (Bayesian, with skopt)

python bayes_tuning.py

• Search space (narrow & practical):

  • lstm_units ∈ [32, 64]

  • dropout_rate ∈ [0.1, 0.3]

  • batch_size ∈ [16, 32]

  • window_size ∈ [20, 40]

• Saves:

  • Best config back into config.yaml (as plain numbers)

  • Full result object → results/bayes_opt_result.pkl

The objective runs short (few epochs, early stopping) for speed. For final training, re-run train.py with the tuned config.

Model Architecture

Input: [window_size, num_features] → LSTM(lstm_units, return_sequences=True) → Dropout(dropout_rate) → LSTM(lstm_units) → Dropout(dropout_rate) → Dense(dense_units=1) # next-day Close Loss: mse | Optimizer: adam

Acknowledgements

• ta technical analysis library

• TensorFlow/Keras, scikit-learn, scikit-optimize