Store Sales Forecasting - Model Comparison Project

This project implements and compares three different approaches for time series forecasting of store sales:

1. Enhanced TCN (Temporal Convolutional Network with residual blocks)
2. SARIMA (Seasonal AutoRegressive Integrated Moving Average)
3. Hybrid (SARIMA + TCN combination)

📁 Project Structure

ML/
├── main.ipynb                          # Enhanced TCN model
├── arima_baseline.ipynb                # SARIMA baseline model
├── hybrid_model.ipynb                  # Hybrid SARIMA+TCN model
├── model_comparison.ipynb              # Comprehensive model comparison
├── data/                               # Training and test data
│   ├── train.csv
│   ├── test.csv
│   ├── oil.csv
│   ├── holidays_events.csv
│   └── stores.csv
└── README.md                           # This file

🚀 Quick Start

1. Install Dependencies

pip install pandas numpy torch scikit-learn matplotlib pmdarima statsmodels joblib seaborn

2. Run Models (in order)

Option A: Run All Models

# 1. Enhanced TCN (fastest, ~10-15 minutes)
jupyter notebook main.ipynb

# 2. SARIMA Baseline (slow, ~30-60 minutes)
jupyter notebook arima_baseline.ipynb

# 3. Hybrid Model (moderate, ~20-30 minutes)
jupyter notebook hybrid_model.ipynb

# 4. Compare All Models
jupyter notebook model_comparison.ipynb

Option B: Quick Start (Just TCN)

# Run only the TCN model for fastest results
jupyter notebook main.ipynb

📊 Model Comparison

Model	Type	Training Time	Strengths	Weaknesses
TCN	Deep Learning	⚡ ~10-15 min	Fast, captures complex patterns, handles all series at once	Black box, requires GPU
SARIMA	Statistical	⏱ ~30-60 min	Interpretable, captures linear trends/seasonality	Slow with many series, linear only
Hybrid	Combined	⚡⏱ ~20-30 min	Best of both worlds, interpretable + powerful	More complex to maintain

🎯 Recommended Workflow

For Best Predictions:

1. Run all three models
2. Use the Weighted Ensemble from model_comparison.ipynb
3. Weights: 30% TCN, 20% SARIMA, 50% Hybrid

For Production Deployment:

1. Use Enhanced TCN from main.ipynb
2. Fast inference, easy to update with new data
3. Scales well with more data

For Stakeholder Presentations:

1. Use SARIMA from arima_baseline.ipynb
2. Interpretable coefficients
3. Clear trend/seasonality decomposition

📈 Key Features

Enhanced TCN (main.ipynb)

• ✅ Residual blocks with skip connections
• ✅ Batch normalization & dropout
• ✅ Learning rate scheduling & early stopping
• ✅ Gradient clipping
• ✅ 180-day lookback (6 months for seasonality)
• ✅ Integrated oil prices & holidays

SARIMA Baseline (arima_baseline.ipynb)

• ✅ Automatic parameter selection (auto_arima)
• ✅ Parallel processing for 1,782 time series
• ✅ Weekly seasonality (m=7)
• ✅ Robust error handling with fallback to mean
• ✅ Validation metrics & visualizations

Hybrid Model (hybrid_model.ipynb)

• ✅ Stage 1: SARIMA captures trend/seasonality
• ✅ Stage 2: TCN models residuals (non-linear patterns)
• ✅ Final prediction = SARIMA + TCN_residuals
• ✅ Best accuracy with interpretability

📋 Output Files

Submissions

• submission.csv - TCN predictions
• submission_sarima.csv - SARIMA predictions
• submission_hybrid.csv - Hybrid predictions
• submission_ensemble_simple.csv - Simple average ensemble
• submission_ensemble_weighted.csv - ⭐ Recommended weighted ensemble

Model Files

• best_model.pth - Best TCN model (validation)
• final_model.pth - Final TCN model (full data)
• sarima_models.pkl - Trained SARIMA models
• hybrid_sarima_models.pkl - Hybrid SARIMA component
• hybrid_tcn_final.pth - Hybrid TCN component

Visualizations

• predictions_visualization.png - TCN sample predictions
• sarima_validation_predictions.png - SARIMA validation
• hybrid_decomposition.png - SARIMA vs residuals
• hybrid_predictions_visualization.png - Hybrid predictions
• model_comparison_*.png - Various comparison plots

🔧 Customization

Adjust TCN Parameters (main.ipynb)

INPUT_LENGTH = 180        # Lookback window (days)
OUTPUT_LENGTH = 16        # Forecast horizon (days)
BATCH_SIZE = 32          # Batch size for training
hidden_channels = 128    # Model capacity
num_blocks = 4           # Number of residual blocks
dropout = 0.2            # Dropout rate

Adjust SARIMA Parameters (arima_baseline.ipynb)

max_p = 3               # Max AR order
max_q = 3               # Max MA order
max_P = 2               # Max seasonal AR order
max_Q = 2               # Max seasonal MA order
m = 7                   # Seasonal period (weekly)

📊 Performance Metrics

The model_comparison.ipynb notebook provides:

• ✅ RMSE, MAE, MAPE on validation set
• ✅ Prediction distribution analysis
• ✅ Time series visualizations
• ✅ Correlation between models
• ✅ Statistical comparison tables

🎓 Key Insights

Why TCN over LSTM/GRU?

• ✅ Parallel processing (faster training)
• ✅ Longer effective history (dilated convolutions)
• ✅ More stable gradients
• ✅ Better for very long sequences

Why SARIMA Still Matters?

• ✅ Interpretable for business stakeholders
• ✅ Works with less data
• ✅ Captures explicit seasonality
• ✅ Good baseline for comparison

Why Hybrid Approach?

• ✅ SARIMA handles predictable patterns
• ✅ TCN handles complex residuals
• ✅ Combines strengths of both
• ✅ Often outperforms either alone

🐛 Troubleshooting

SARIMA Takes Too Long

• Reduce the number of series (sample first)
• Simplify parameter space in auto_arima
• Use more CPU cores (n_jobs=-1)

Out of Memory (TCN)

• Reduce BATCH_SIZE
• Reduce INPUT_LENGTH
• Reduce hidden_channels
• Use CPU instead of GPU if available

Poor Predictions

• Check data quality and missing values
• Increase INPUT_LENGTH for more context
• Try different train/val split ratios
• Experiment with different model architectures

📚 References

• TCN Paper: An Empirical Evaluation of Generic Convolutional and Recurrent Networks for Sequence Modeling
• SARIMA: Box, Jenkins, Reinsel - Time Series Analysis
• Hybrid Approaches: Zhang (2003) - Time series forecasting using a hybrid ARIMA and neural network model

📞 Support

For questions or issues:

1. Check the troubleshooting section
2. Review the notebook comments
3. Examine the visualization outputs
4. Compare with baseline metrics

🎉 Results

After running all notebooks, you'll have:

• ✅ 5 different submission files
• ✅ Multiple trained models
• ✅ Comprehensive visualizations
• ✅ Performance metrics comparison
• ✅ Clear recommendation for production

Recommended submission: submission_ensemble_weighted.csv

---

Happy Forecasting! 📈🚀