This study investigates whether model complexity necessarily correlates with forecasting performance by comparing the DLinear, which is structurally simple, against Informer and ARIMA-hybrid models. To validate this, we utilized simulation data reflec...
This study investigates whether model complexity necessarily correlates with forecasting performance by comparing the DLinear, which is structurally simple, against Informer and ARIMA-hybrid models. To validate this, we utilized simulation data reflecting seven distinct statistical properties, as well as real-world datasets. The experimental results demonstrate that DLinear outperformed complex deep learning and hybrid models in real-world data analysis. However, the effectiveness of hybrid models was partially observed in specific simulation scenarios where non-linearity was dominant. In conclusion, this study suggests that it is critical to identify the characteristics of the target data and select an appropriate model accordingly. Furthermore, it implies that in certain cases, a parsimonious approach—structurally decomposing the data and employing a direct forecasting strategy—can be a more efficient alternative.