<![CDATA[Purpose: This study evaluates the performance of distributed lag, autoencoder, and LSTM autoencoder methods in detecting point anomalies in simulated data generated from Jakarta's Air Quality Index (AQI). The evaluation is conducted across several simulation scenarios that represent factors that may influence anomaly detection performance. Methods: Simulation scenarios were constructed by varying two anomaly characteristics: anomaly percentage (0.3%, 0.5%, and 1.0%) and anomaly depth (4.4σ, 4.7σ, and 5.0σ), yielding 90 datasets generated via repeated experiments. Anomaly detection was performed using a forecasting-based approach with a 4σ threshold on prediction errors. Model performance was evaluated using mean absolute percentage error (MAPE) for forecasting accuracy and balanced accuracy for anomaly detection. Result: Increasing anomaly percentage significantly degrades both forecasting and anomaly detection performance across all methods. In contrast, anomaly depth has no significant effect on forecasting accuracy but strongly influences detection performance. Among the evaluated methods, the distributed lag model consistently shows the most robust anomaly detection performance across scenarios, outperforming the autoencoder and LSTM autoencoder, particularly at higher anomaly percentages and depths. Novelty: This study introduces a more rigorous and structured evaluation framework for anomaly detection by integrating three key contributions. First, it provides a unified comparison of distributed lag, autoencoder, and LSTM autoencoder methods within a single experimental setting, a limitation in prior studies. Second, it employs a controlled simulation design that systematically varies the anomaly percentage and depth while preserving key characteristics of empirical AQI data, enabling a more objective assessment of model performance across diverse anomaly conditions. Third, it uses ANOVA and interaction analysis to formally examine the effects of anomaly characteristics on both forecasting and detection performance, moving beyond purely descriptive comparisons commonly used in previous research.]]>
