PR: LSTM Time Series Forecasting in R

This PR introduces a fully documented implementation of a Long Short-Term Memory (LSTM) neural network in R for time series prediction and forecasting.
The implementation leverages the keras and tensorflow packages to build, train, and evaluate LSTM models capable of learning temporal dependencies in sequential data.

Overview

The provided LSTM workflow covers:

• Data Preprocessing:

  • Normalization of time series data to [0,1] range
  • Sequence creation for supervised learning (input sequences and target values)

• Model Architecture:

  • Single-layer LSTM network with configurable units and dropout
  • Dense output layer for regression tasks
  • Compilation with Adam optimizer and Mean Squared Error loss

• Training and Evaluation:

  • Training with configurable epochs and batch size
  • Validation split and monitoring of training history
  • Evaluation on test data with metrics: MSE, RMSE, MAE, and R-squared

• Prediction Capabilities:

  • Single-step and multi-step ahead predictions
  • Denormalization of predictions for interpretation
  • Visualization of actual vs. predicted series using ggplot2

• Best Practices:

  • Sequence length selection based on temporal dependencies
  • Dropout regularization to prevent overfitting
  • Early stopping, model checkpointing, and hyperparameter tuning

This LSTM implementation is suitable for:

• Synthetic or real-world time series (e.g., sine waves, stock prices, sensor data)
• Forecasting applications requiring learning of sequential patterns
• Multi-step ahead prediction and analysis of temporal trends

Complexity

• Training Complexity: O(n_samples × seq_length × lstm_units) per epoch
• Inference Complexity: O(n_samples × seq_length × lstm_units)

The approach demonstrates how LSTMs can model sequential dependencies more effectively than traditional regression or moving-average methods.