Tata Power Next-Day Close Price Prediction

This project trains and tests an LSTM model to predict next-day closing prices for equities using daily OHLCV data and technical indicators.

The pipeline is generic for any company/stock as long as you provide a similar daily CSV (with Date, Open, High, Low, Close/Price, Volume). In this repository we have tested it on Tata stocks (Tata Power / Tata Motors) and obtained good results.

Files

• v3_delta.py
  Trains the LSTM model on TataPower_2005_2025.csv using:

  • Window size: 14 previous trading days
  • Features: Price, Open, High, Low, Vol., Volatility, RSI, Momentum
  • Target: next-day log return of Price, later converted back to price.
  • Outputs:
    • v3_delta_savedmodel/ (TensorFlow SavedModel)
    • v3_delta_scaler_X.pkl (StandardScaler for features)
    • v3_delta_scaler_y.pkl (RobustScaler for target)

• test_range_v3_delta_csv.py
  Loads the saved model and scalers, then evaluates predictions over a chosen date range (currently 2025-11-01 .. 2025-12-31) using the CSV file. It prints:

  • Per-day actual vs predicted next-day close
  • MAE, RMSE, and % within ±50 points
  • How many days the model predicted a drop (negative log return)
  • A matplotlib plot of actual vs predicted next-day close.

• TataPower_2005_2025.csv
  Historical Tata Power price data used for training and testing.

• v3_delta_savedmodel/
  Trained TensorFlow model directory.

• v3_delta_scaler_X.pkl, v3_delta_scaler_y.pkl
  Scalers used for input features and target, saved after training.

• Figure_1.png
  Example plot of Actual vs Predicted next-day Close from a test run.

Example Figure

Below is an embedded example plot from the project:

How to Run

1. Environment

   • Python 3.11
   • TensorFlow 2.15.x
   • Required Python packages: tensorflow, keras, numpy, pandas, scikit-learn, matplotlib.

2. Train / Retrain the Model

python v3_delta.py

This:

• Reads TataPower_2005_2025.csv
• Trains the LSTM with a 14-day window
• Saves the model and scalers.

3. Run the Test Script

python test_range_v3_delta_csv.py

This:

• Loads v3_delta_savedmodel/ and the scalers
• Evaluates from 2025-11-01 to 2025-12-31
• Prints metrics and shows the Actual vs Predicted plot
• Prints how many days the model predicted a price drop.

Notes

• The main goal is to keep next-day close predictions within ±50 points of the actual close.
• The shorter 14-day window makes the model more sensitive to recent market moves and better at predicting both up and down days.