Time series forecasting is of paramount importance for the daily operations of various enterprises. Particularly, addressing time series problems with short sequence lengths presents unique challenges, as product life cycles continue to decrease, but conventional time series models still require adequate sample sizes. In this study, we introduce a novel stacked generalization method that incorporates the self-attention mechanism, allowing for the effective utilization of diverse models in short time series forecasting. Through the application of real data, we demonstrate the remarkable potential of our proposed method and show its ability to overcome the limitations of traditional approaches in tackling short time series forecasting problems.