Time-series are widely used for representing non-stationary data such as weather information, health related data, economic and stock market indexes. Many statistical methods and traditional machine learning techniques are commonly used for forecasting time series. With the development of deep learning in artificial intelligence, many researchers have adopted new models from artificial neural networks for forecasting time series. However, poor performance of applying deep learning models in short time series hinders the accuracy in time series forecasting. In this paper, we propose a novel approach to alleviate this problem based on transfer learning. Existing work on transfer learning uses extracted features from a source dataset for prediction task in a target dataset. In this paper, we propose a new training strategy for time-series transfer learning with two source datasets that outperform existing approaches. The effectiveness of our approach is evaluated on financial time series extracted from stock markets. Experiment results show that transfer learning based on 2 data sets is superior than other base-line methods.