Forecaster

A modular time series forecasting framework that integrates symbolic regression, ARX models, and attention-based feature selection. Designed for multi-step forecasting and model evaluation.

Features

• AR and ARX models with customizable lag structures
• Symbolic regression with nonlinear transformations
• Attention-based feature selection using PyTorch
• Rolling window validation and quality of forecast metrics

Getting Started

1. Git clone the repository

git clone https://github.com/Youseffekri/Forecaster.git

cd Forecaster

2. Install dependencies:

pip install -r requirements.txt

3. Run the example script:

python Examples\Example_Covid19.py

Models

• AR_YW: Autoregressive model using Yule-Walker estimation
• ARX, ARX_D: AR model with exogenous inputs
• ARX_Symb, ARX_Symb_D: Symbolic regression-based ARX
• MHAttnRegressor: Multi-head attention for feature selection

Evaluation Metrics

• Mean Squared Error (MSE)
• Mean Absolute Error (MAE)
• Coefficient of Determination (R² and Adjusted R²)
• Symmetric Mean Absolute Percentage Error (SMAPE)

License

MIT License

Author

Yousef Fekri Dabanloo
March 2025