Sparse Dynamic Factor Modeling of Some Selected Climatic Variables

The classical forecasting models struggle to handle missing data, a common issue in climate data, due to irregular reporting intervals or sensor failures. Incomplete datasets can lead to biased or unreliable forecasts, further complicating efforts to predict climatic variables accurately. This study aims to examine the performance of the Sparse dynamic factor models on climate data. Its performance is compared with classical models, such as ARIMA, PCA, Two-Stage DFM, EM-Based DFM, Sparse DFM, Lasso, and Group Lasso. The study integrates a traditional statistical approach with penalized likelihood optimization, ensuring the inclusion of sparse, interpretable models. The dataset employed in this study was extracted from the Nigerian Meteorological Agency (NiMet) and the National Bureau of Statistics (NBS) statistical bulletin 2023. The data includes Annual Average Mean Surface Air Temperature, Annual Precipitation, Number of Days with Heat Index > 35°C, and Maximum Number of Consecutive Wet Days. The findings of the study revealed that group Lasso consistently yielded the lowest MSE across key variables, Air Temperature (MSE = 0.3854), Precipitation (921.27), Heat Days (296.85), and Wet Days (748.90) outperforming all benchmark models. Results also showed that, ARIMA, PCA, and Two-Stage DFM recorded substantially higher errors, highlighting their inability to capture intricate, nonlinear dependencies present in climate processes.