Comparing Machine Learning Methods for Offline Applications with Short Lookback and Limited Input

This study focuses on a comparative study of machine learning methods on offline weather forecasting with short lookback windows and limited computational resources. Using 90 days of single-station GSOD inputs, models predict 15-day horizons for temperature, precipitation, wind speed, and visibility. Evaluation with NashSutcliffe efficiency, RMSE, and inference time shows that Linear Regression is a surprisingly strong and stable baseline, excelling in wind speed and remaining competitive across variables. Transformer models perform best for temperature by capturing long-range dependencies, while sequence-to-sequence GRUs outperform others on precipitation and visibility. In contrast, XGBoost and persistence baselines consistently underperform in this constrained setup. In inference time, LR outperform all other methods due to its simplicity. The results indicate that simple linear models can excel in this scenario compare to deep learning approaches, while specialized neural architectures provide targeted gains, suggesting a combination of models could be the most effective for practical low-resource forecasting.