Time Series Forecasting Benchmark

This repository presents a consolidated benchmarking study comparing classical, machine learning, and deep learning models for time-series forecasting under limited data conditions.

Models Evaluated

• ARIMA (statistical baseline)
• XGBoost (tree-based regression with lag features)
• LSTM (recurrent neural network)
• Transformer (attention-based deep learning model)

Methodology

• Time-series data is split into training, validation, and test sets.
• Hyperparameter tuning is performed for LSTM and Transformer models using validation RMSE.
• Final evaluation is conducted on a held-out test set using RMSE as the primary metric.
• The best-performing model is used for long-horizon forecasting.

Key Findings

• Deep learning models outperform traditional approaches on this dataset.
• LSTM achieves the lowest test RMSE, showing better generalisation in low-data settings.
• Transformer performs well during validation but shows mild overfitting on the test set.

Notes

• Results should be interpreted with caution due to the small test size.
• Long-horizon forecasts are generated using a recursive strategy, which may accumulate error.

Requirements

• Python 3.x
• NumPy
• Pandas
• Scikit-learn
• Statsmodels
• PyTorch
• XGBoost

Usage

Run the notebook sequentially to reproduce preprocessing, model training, evaluation, and forecasting results.