Etsy

This is my time series forecasting system leveraging ARIMA/SARIMA, Prophet, LSTM networks and ensemble methods along with automated feature engineering for real-time predictive analytics on simulated streaming data. The project includes automated feature engineering (lag features, rolling statistics, date-time components) and simulates real-time streaming data ingestion to update forecasts in near real-time.

The project was a part of my freelance work, published with permission. For privacy purposes, data was removed.

What's inside:

• Time series models
  ARIMA/SARIMA using statsmodels, Prophet forecasting and LSTM networks built with TensorFlow/Keras.
• Feature engineering
  Automated generation of lag features, rolling averages and date-time components.
• Ensemble forecasting
  Combination of classical and ML-based forecasts using stacking ensembles.
• Real-time simulation
  Demonstrates near real-time forecasting with simulated streaming data.
• Evaluation metrics
  MAE, RMSE, MAPE and custom time-series cross-validation routines.
• Modular code
  Clean, modular and well-documented code following best practices and designed for production deployment.

Installation (clone, navigate, create a venv, install dependencies):

git clone https://github.com/avrtt/etsy.git
cd etsy
python -m venv venv
source venv/bin/activate
pip install -r requirements.txt

Run the main script to execute the entire pipeline:

python src/main.py

The main script will:

• Generate or load synthetic time series data
• Apply feature engineering
• Train multiple forecasting models
• Evaluate each model using various metrics
• Simulate streaming data updates and display real-time forecast updates

Project structure:

• data/ contains the synthetic data (generated automatically if not provided)
• src/ contains the Python modules
  • data_loader.py: data ingestion and synthetic data generation
  • feature_engineering.py: automated feature generation routines
  • evaluation.py: functions for calculating forecast evaluation metrics
  • streaming.py implements simulated streaming data forecasting
  • main.py orchestrates the complete forecasting pipeline
  • models/ contains implementations for:
    • arima_model.py: ARIMA/SARIMA-based forecasting
    • prophet_model.py: forecasting using Prophet
    • lstm_model.py: LSTM network for deep learning forecasts
    • ensemble_model.py: ensemble method combining multiple forecasts

Contributions, issues and feature requests are welcome. Apache 2.0 license.