Canton-level crime risk index for Ecuador
A composite crime risk index covering every canton in Ecuador, combined with KMeans segmentation and a national forecast benchmarked across six time-series models.
The question
How can crime risk be compared fairly across cantons when some have millions of residents and others only a few thousand? And once that measure exists, which model best forecasts the national trend?
Approach
1. Canton-level index. Built from Ministry of the Interior open data (homicides, arrests and seized weapons) and INEC population projections:
- Rates per 10,000 inhabitants.
- Bayesian shrinkage, so that small cantons don’t look extreme because of random noise: each rate is pulled toward the national mean in proportion to how little data backs it.
- Indicator weights derived from PCA, combined into a weighted z-score index.
- A robustness check comparing the index against unadjusted rates and raw counts.
2. Segmentation. KMeans, with the number of clusters chosen using the elbow method, silhouette, Calinski-Harabasz and Davies-Bouldin scores. The result was 6 clusters of cantons with distinct risk profiles.
3. Forecasting. A monthly national index, trained on 2019-2023 and evaluated on 2024-2025, with no information from the test period leaking into training.
Results (test set, 2024-2025)
| Model | RMSE | MAE |
|---|---|---|
| Ensemble | 0.0871 | 0.0679 |
| LSTM | 0.1169 | 0.0963 |
| XGBoost | 0.1340 | 0.1102 |
| ETS | 0.1361 | 0.1221 |
| Prophet | 0.1501 | 0.1362 |
| ARIMA | 0.1609 | 0.1486 |
Stack
pandas, scikit-learn, TensorFlow/Keras, XGBoost, statsmodels, pmdarima and Prophet.