<![CDATA[This study investigates the impact of Dynamic Time Warping (DTW)-based clustering on forecasting accuracy in anomalous time series data. Monthly export values from 30 provinces in Indonesia are clustered using K-Medoids and Fuzzy C-Means (FCM) based on DTW distances reduced through Multidimensional Scaling (MDS). Cluster validation results indicate that FCM demonstrates better stability under anomalous conditions, with a Silhouette Score of 0.62 and a Davies–Bouldin Index of 0.57, compared to K-Medoids with a Silhouette Score of 0.61 and a Davies–Bouldin Index of 0.58. Forecasting performance is systematically evaluated through an expanding-window scheme using standalone ARIMA and Temporal Convolutional Network (TCN) models as baselines against a hybrid ARIMA–TCN approach. The results show that the hybrid ARIMA–TCN model achieves the lowest Mean Absolute Percentage Error (MAPE) in clusters with relatively stable patterns by effectively combining linear and non-linear components. However, for clusters characterized by higher volatility and irregular patterns, the standalone TCN model yields better forecasting accuracy. Furthermore, FCM clustering produces better overall forecasting accuracy, with an average MAPE of 7.11%, compared to 8.42% for K-Medoids. The relatively small difference in evaluation results between clean and empirical data suggests that the proposed DTW–MDS–clustering–forecasting framework maintains consistent performance in the presence of anomalies. The final model is then applied to generate export forecasts for the next 12 periods.]]>
Copyrights © 2026
