Advanced time series forecasting analysis comparing ARIMA, LSTM, and Prophet algorithms for predicting chili prices across 5 major traditional markets in Medan, Indonesia.

• ✅ RMSE: 11,933
• ✅ MAPE: 13.76% (Excellent category)
• ✅ Performance: 66% better than ARIMA, 80% better than Prophet
• ✅ Consistency: Best in all 5 markets

1. Deep learning essential for volatile commodity forecasting (CV=40%)
2. Holiday features NOT needed for LSTM (already captures patterns)
3. ARIMA struggles with non-linear price movements (MAPE 41%)
4. Prophet fails on volatile commodities (MAPE 70% - designed for smooth business metrics)

# Navigate to notebooks and execute in order:
cd notebooks/
jupyter notebook 01_data_cleaning_and_eda.ipynb
jupyter notebook 02_arima_modeling.ipynb
jupyter notebook 03_lstm_modeling.ipynb
jupyter notebook 04_prophet_modeling.ipynb
jupyter notebook 05_model_comparison_and_inference.ipynb

• 📄 Full Report: LAPORAN_FINAL.md (8,500 words)
• 📊 Comparison: results/metrics/model_comparison.csv
• 📈 Visualizations: results/plots/

Winner: LSTM (baseline) - 66% better than ARIMA, 80% better than Prophet

See LAPORAN_FINAL.md for complete:

• Methodology
• Data analysis
• Model architectures
• Statistical tests
• Business recommendations