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
- Deep learning essential for volatile commodity forecasting (CV=40%)
- Holiday features NOT needed for LSTM (already captures patterns)
- ARIMA struggles with non-linear price movements (MAPE 41%)
- 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/
| Rank | Model | RMSE | MAPE (%) | Category |
|---|---|---|---|---|
| 1 | LSTM (baseline) | 11,933 | 13.76 | Excellent ⭐ |
| 2 | LSTM + Holiday | 14,498 | 18.02 | Good |
| 3 | ARIMA | 35,197 | 41.21 | Poor |
| 4 | Prophet + Holiday | 49,684 | 69.94 | Very Poor |
| 5 | Prophet (baseline) | 51,090 | 73.90 | Very Poor |
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