Abstract How can we efficiently determine meta-parameter values for deep learning-based time-series forecasting given a time-series dataset? This paper introduces Xtune, an efficient and novel meta-parameter tuning method for deep learning-based time-series forecasting, leveraging explainable AI techniques. In particular, this study focuses on optimizing the window size for time-series forecasting. Xtune determines the optimal meta-parameter value for these methods and can also be applied to tune the window size for anomaly detection methods that utilize deep learning-based time-series forecasting. Extensive experiments on real-world datasets and forecasting methods demonstrate that Xtune efficiently identifies the optimal meta-parameter value and consistently outperforms the existing methods in terms of execution speed.