<![CDATA[In predictive modeling, data partitioning mainly involves dividing the dataset into a training set for learning model parameters and a testing set for assessing generalizability through predictive performance. In time series forecasting, non-random data splitting is commonly used where initial observations are used for training, a subsequent portion for validation and latest observations are utilized for testing. However, the effectiveness of random data partitioning where observations are randomly distributed into training, validation and testing subsets remains underexplored within the context of time series data. This study investigates the suitability of random data partitioning in time series forecasting using both stationary and non-stationary simulated datasets, as well as real-world data. Feedforward neural network (FFNN) and Support vector regression (SVR) models were implemented, with model performance optimized through systematic trial-and-error hyperparameter tuning. Under random data-splitting approach, 30 different training and testing subsets are generated to assess the stability and robustness of model performance across different sample compositions. Random data partitioning is implemented at the level of constructed supervised learning instances, where each instance consisted of lagged input variables and their corresponding target values, forming distinct input–output pairs. The allocation of these instances to training and testing subsets is carried out entirely at random, without preserving sequential ordering or grouping temporally adjacent observations, while strictly maintaining original input–output correspondence within each constructed instance. The findings indicate that, for both simulated and real-world datasets, random data splitting resulted in improved predictive performance of both models, yielding lower error metrics compared to non-random splitting. These results suggested that random data splitting can enhance generalization and forecasting performance in time series applications with appropriate models. The study provides valuable empirical evidence on data partitioning strategies, supporting researchers and practitioners in making more informed decisions regarding model evaluation and selection in time series forecasting tasks.]]>
